---
url: "https://questionstar.com/blog/social-desirability-bias/"
title: "Social desirability bias: how to ask sensitive questions"
description: "What social desirability bias is, which survey questions are sensitive, and how anonymity, careful wording and randomized response lead to more honest answers."
lastmod: "2026-11-11"
type: "blog"
---

# Social desirability bias: how to ask sensitive questions

*Why don't respondents always say what they think or do, and how do I ask sensitive questions so that I still get honest answers?*

What social desirability bias is, which survey questions are sensitive, and how anonymity, careful wording and randomized response lead to more honest answers.

Social desirability bias is the tendency of respondents to shape their answers around what's considered acceptable in their social environment rather than around what's actually true (Möhring and Schlütz 2019, p. 60). You can't eliminate it completely, but you can reduce it (ibid., p. 64). Use a self-administered survey with credible anonymity, word questions without putting respondents on the spot, and turn to indirect techniques such as randomized response only when necessary.

Let's go through it step by step. You'll learn:

- [what social desirability bias is](#definition) and why it isn't automatically a deliberate lie,
- [which questions are sensitive](#sensitive-questions), from income to job satisfaction,
- [what the survey mode](#mode) and [anonymity or confidentiality](#anonymity) can do,
- [how to word sensitive questions](#wording) without manipulating anyone,
- [what randomized response, the item count technique and the crosswise model can do](#indirect-techniques), and where their limits are.

At the end, you'll find an [employee survey](#employee-survey) example, a checklist and answers to frequently asked questions.

I come across this question in many forms. The student who wants to ask about alcohol consumption for her bachelor's thesis and suspects that nobody will tick "more than ten drinks a week". The HR manager whose employee survey shows surprisingly good ratings for every single manager. The client in whose survey almost everyone donates "regularly". In all three cases, social desirability bias may play a role, and for all three there are remedies worth knowing before the questionnaire goes online.

## What is social desirability bias?

Möhring and Schlütz define socially desirable responding as the tendency of respondents to align their answers with what is regarded as socially accepted and desirable within the normative system of their social world, which distorts the "true value" (2019, p. 60). In German-language research, the term is "soziale Erwünschtheit". In English, it's usually called social desirability bias, SDB for short.

The bias runs in two directions. Behavior considered good tends to be exaggerated (overreporting). Behavior considered bad tends to be played down (underreporting) (Krumpal 2013, cited in Möhring and Schlütz 2019, p. 60). Among the topics that tend to be underreported, Bradburn et al. list illness, crime, tax evasion, drug use and alcohol consumption. Topics that tend to be overreported include voting, reading, attending cultural events, continuing education and donating to charity (Bradburn et al. 2004, cited in Schnell 2019, pp. 44-45).

This matters for how you read your data: a socially desirable answer isn't automatically a deliberate lie. People can embellish their behavior without being aware of it, because they've internalized the values of their environment. Alongside this, there's deliberate impression management (Möhring and Schlütz 2019, p. 61). Schnell even describes the socially desirable answer as an often rational action: when an honest answer could put the respondent's other goals at risk, for example with sensitive questions or when others are listening, the embellished answer makes sense from the respondent's point of view (2019, p. 44).

This leads to the most important idea in this article:

> **You don't get honest answers by appealing to people's honesty, but by making the honest answer safe.**

One more subtlety: what's desirable depends on the reference group. In some groups, behavior that's otherwise frowned upon is considered desirable and is then exaggerated (Schnell 2019, p. 45). In one US example, teenagers reported considerably more experience with cannabis to an interviewer of their own age than to an older female interviewer (Jacob et al. 2019, p. 54).

## Which survey questions are sensitive?

According to Jacob et al., sensitive topics are those respondents are reluctant to discuss because they consider them too private or embarrassing, because they involve socially undesirable attitudes or behavior, or because they fear negative consequences from taking part in the survey (2019, p. 174). Sensitive questions more often lead to refusals, false answers and breakoffs (ibid.).

It helps to distinguish three meanings of sensitivity: intrusion into privacy, risk of disclosure to third parties and social desirability (Tourangeau et al. 2000, cited in Schnell 2019, p. 44). Each calls for a different remedy. Explaining the purpose helps address concerns about intrusion into privacy. Genuine anonymity helps address fear of disclosure. Neutral wording that doesn't devalue any answer helps reduce the pressure to give a socially desirable answer.

| Topic | Typical bias | What helps | Source |
|------|------|------|------|
| Income | Refusals and understated amounts | Income brackets instead of an exact figure, question at the end | Jacob et al. 2019, p. 175; Schnell 2019, p. 109 |
| Health, illness | Underreporting | Explain the purpose, anonymity, self-completion | Bradburn et al. 2004, cited in Schnell 2019, p. 44 |
| Alcohol, drugs | Underreporting, overreporting in some groups | Anonymity, neutral wording, an indirect technique if needed | Schnell 2019, pp. 44-45; Jacob et al. 2019, p. 54 |
| Donations, volunteering | Overreporting | A specific time period, neutral wording | Bradburn et al. 2004, cited in Schnell 2019, p. 45 |
| Voting behavior, turnout | Overreporting of turnout, and in Germany voting behavior is considered particularly sensitive | Self-completion, anonymity | Krumpal 2013, cited in Möhring and Schlütz 2019, p. 60; Schnell 2019, p. 44 |
| Job satisfaction, ratings of managers (in surveys run by the employer) | Answers shaped by what management should believe, fear of being identified | Anonymity, minimum group size in reports, no pressure | Linke 2018, p. 8; Wiertalla 2013, p. 102 |

By the way, what counts as sensitive changes over time and with the target group. According to Jacob et al., questions about the number of sexual partners are hardly considered sensitive anymore and lead to only a few refusals, whereas questions about income still are (2019, pp. 174-175). If in doubt, check it in the pretest (ibid., p. 174).

An example shows how big the difference can be. In the German ZUMA Social Science Bus II survey of 1990, about 55% of respondents refused to answer the direct question about their income. A follow-up question with income brackets reduced this share to about 16% (Jacob et al. 2019, p. 175). That's still a high refusal rate, but the follow-up question recovered many responses that would otherwise have been missing.

## Online or with an interviewer: the survey mode

One important lever lies not in the individual question, but in the situation in which it's answered. Möhring and Schlütz summarize the state of research as follows: in self-administered surveys, by mail or online, the effect of social desirability is smaller than in telephone or face-to-face interviews, where an interviewer is present (Möhring and Schlütz 2019, p. 62, with further references). Schnell also usually finds the largest differences between survey modes when comparing modes with and without an interviewer, but he stresses that the differences are rather small overall (2019, p. 308).

So for sensitive topics, an online survey has a real advantage: nobody sees your facial expression or hears the hesitation in your voice. Even face-to-face interviews make use of this, by having respondents fill in the sensitive parts themselves and return them in a sealed envelope (Jacob et al. 2019, p. 117).

Two caveats belong here. First, the online survey replaces the interviewer with a different worry: who will see my answer later? More on this in a moment. Second, the setting matters. Whether other people are present influences the answers (Schnell 2019, p. 45), although family members tend to act as a check and dampen embellishment (Krumpal 2013, cited in Möhring and Schlütz 2019, p. 62). Someone filling in a survey at a shared computer isn't alone.

The levers you control can be arranged in three levels:

The order is intentional. The levels at the bottom cost little and affect all questions. The indirect techniques at the top are expensive and are only worth it for individual, truly sensitive questions.

## Anonymous or confidential? A difference that matters

Möhring and Schlütz advise that you should always assure respondents of anonymity in the questionnaire and the instructions, to defuse the survey situation (2019, p. 62). That's right, but there's a condition: the assurance must be true. And this is exactly where two words are often confused.

> **Anonymous** means: the answers can't be linked to any person, not even by you.
> **Confidential** means: a link would be technically possible, but you promise not to make it and to protect the data.

Both are legitimate. What's dishonest is writing "anonymous" when you mean "confidential". With personal links, there's one more subtlety: the system knows who has already taken part. Otherwise, it couldn't send targeted reminders. In that case, the answers are anonymous, but the participation status is confidential.

Schnell is strict here: as soon as access to a web survey is controlled by a username, a password or a personal code, the respondent can in principle be identified. Respondents then have to believe the assurances of anonymity or confidentiality, and within organizations in particular, for example in employee surveys, their mistrust can make the survey useless (2019, p. 296). In a footnote, he adds that within an organization, the IP number usually reveals which computer an answer came from. For topics with criminal relevance, he notes, neither the operators' servers nor the internet providers' logs are safe from seizure. Strictly speaking, he concludes, a web survey can hardly be considered "anonymous" (ibid., n. 36).

That's why I recommend three things.

**Say exactly what happens.** For surveys with an ID number, Schnell recommends being open about it: explain in the invitation what the number is for and that it will be separated from the answers once the responses are in (2019, pp. 244-245). For an online survey with a personal link, this could read:

> ℹ️
> 
> We're sending you a personal survey link so we can send reminders only to people who haven't taken part yet. Your answers are stored separately from your name and email address. Results are reported only for groups of [number] people or more.

Agree on the text with your data protection officer beforehand.

Only write a sentence like this if it's true for your survey. You're responsible for setting the minimum group size and keeping to it: QUESTIONSTAR doesn't automatically hide small groups in filtered reports.

**Protect small groups.** If you ask for department, age group, gender and length of service, you can link almost every answer to a person in a team of seven without ever seeing a name (see Wiertalla 2013, p. 102). So ask only for the characteristics you'll actually analyze, in categories that are as broad as possible.

**Only promise what you can deliver.** According to the literature on employee surveys, anonymity also means that nobody finds out who took part at all, and that neither taking part nor not taking part leads to disadvantages (Wiertalla 2013, p. 185, with further references). How to agree on this with the works council and what minimum group size makes sense for reports is covered in the article [Employee survey](/blog/employee-survey/#anonymity). The legal side, that is, when a survey counts as anonymous or pseudonymized under the GDPR, is covered in the article [Data protection in online surveys](/blog/data-protection-in-online-surveys/#when-is-the-survey-anonymous-and-what-is-pseudonymization).

By the way, long explanations about anonymization tend to raise concerns rather than ease them, as I described in the article [Questionnaire design](/blog/questionnaire-design/#structure). One clear, true sentence is better than a paragraph full of assurances.

## How to word sensitive survey questions: what helps and what goes too far

Once you've created the right conditions, focus on the question itself. Malhotra and Birks provide a classic list of six techniques: place sensitive topics at the end, "preface the question with a statement that the behaviour of interest is common", ask in the third person, hide the question among questions respondents are willing to answer, provide response categories instead of asking for specific figures, and use randomized techniques (2007, p. 380). Not all of these rules are equally good. Let's look at them one at a time.

**Explain the purpose.** Respondents are reluctant to give information if they see no legitimate reason for it. A sentence like "To understand how needs differ between age groups, we need to know your age" makes the question understandable (Malhotra and Birks 2007, p. 380; [Marx 2026, slide 106](/principles-of-survey-research/chapter-3/slide-106.html)). This works mainly against the feeling that privacy is being invaded.

**Ranges instead of exact figures.** This is the rule with the most tangible evidence, see the income example above. Instead of asking for exact net income, offer income brackets, for example in increments of 500 euros, with an open-ended top bracket (Jacob et al. 2019, p. 175). Response ranges can also help when asking about alcohol consumption, working hours or days of sick leave. The ranges should cover the whole spectrum, and the answer options should be balanced (ibid., p. 136).

**Ask neutrally, don't devalue any answer.** Leading questions such as "Surely you also think that ..." are off-limits (Jacob et al. 2019, p. 136). With sensitive behavior, the issue is the mirror image: the question mustn't mark the honest answer as embarrassing. "Did you fail to vote?" turns a yes into an admission of failure. "In the last national election, did you vote or not?" doesn't devalue either answer.

**Normalize, but honestly.** Jacob et al. allow one exception to the ban on leading questions: if you want to measure how common socially undesirable behavior is, it can make sense to use wording that softens the behavior or points out that it's common. But it mustn't be mistaken for the researcher's own opinion, and it should refer to a group that isn't taking part in the survey, such as "the majority of the population" (2019, p. 136). This is also what the second technique in Malhotra and Birks means, with a sentence such as "Recent studies show that most European consumers are in debt" (2007, p. 380). Here I draw a clear line: a normalizing sentence has to be true. If you write something like this, you need a source for it. Otherwise, you're manipulating your respondents with an invented norm. Usually a neutral introduction that announces the topic and doesn't judge any answer is enough:

> ℹ️
> 
> The next question is about alcohol. Every answer is equally important to us. On how many of the last seven days did you drink alcohol?
> 
> (0 days / 1 to 2 days / 3 to 4 days / 5 to 7 days / prefer not to answer)

**Third person and "other people": best avoided.** The rule of asking in the third person appears in many textbooks, including my course ([Marx 2026, slide 108](/principles-of-survey-research/chapter-3/slide-108.html)). Möhring and Schlütz, however, consider shifting the behavior onto others and appealing to the bandwagon effect undesirable from a research ethics point of view (2019, p. 63). They also report that data quality is higher when respondents are addressed directly rather than asked indirectly about "some" or "other" people (Yeager and Krosnick 2012, cited in Möhring and Schlütz 2019, p. 63). That's why the example above has no introduction along the lines of "Some people ..., other people ...". My advice: ask directly, but in a situation where the direct answer is safe.

**An opt-out answer, but the right one.** Schnell distinguishes, among other things, between "don't know" and a refusal to answer, and recommends recording the two separately right from data collection (2019, p. 46). For sensitive questions, "prefer not to answer" is the appropriate opt-out. "Can't judge" only belongs to questions where someone may genuinely lack the knowledge, for example how well another department works. Respondents generally know about their own alcohol consumption, though they may not remember the exact amount. Response ranges and a short reference period can help with recall. Keep "don't know" and "prefer not to answer" separate.

**The order.** Most sources recommend putting sensitive questions at the end. By then, trust has developed, and if someone breaks off, you at least have the other answers (Schnell 2019, p. 109; Malhotra and Birks 2007, p. 380; [Marx 2026, slide 110](/principles-of-survey-research/chapter-3/slide-110.html)). Weinreich and von Lindern see this differently for customer surveys and consider the middle of the questionnaire the best place (2008, p. 64). I think the end is the safer choice, especially for income. The sources agree on the funnel: first general questions on the topic, then the specific ones (Schnell 2019, p. 34; Jacob et al. 2019, p. 179). How order shifts answers in other ways is described in the article [Order effects in surveys](/blog/order-effects-in-surveys/).

**Check it in the pretest.** Ask pretest participants to rate how strongly an item invites a desirable answer. After data collection, a conspicuously high number of skipped answers can point to sensitive questions (Möhring and Schlütz 2019, p. 62).

Here's what this looks like in three examples:

| Instead of | Better | Why |
|------|------|------|
| "What is your monthly net income? \_\_\_\_ euros" | "Which bracket does your monthly net income fall into?" with brackets and "prefer not to answer" | Brackets reduce refusals (Jacob et al. 2019, p. 175) |
| "Do you donate to good causes regularly?" | "In the last twelve months, have you donated money to a charitable organization? (yes / no / prefer not to answer)" | A specific time period instead of self-image, both answers neutral |
| "Are you satisfied with your manager?" (yes / no) | "How satisfied are you with the support you get from your manager?" (very dissatisfied / somewhat dissatisfied / neither satisfied nor dissatisfied / somewhat satisfied / very satisfied) | A symmetric scale with a midpoint instead of a yes-or-no verdict on your own manager |

The third question deliberately has no "can't judge": everyone knows how satisfied they are themselves.

## Indirect techniques: randomized response, item count, crosswise

Some questions remain sensitive no matter how well you word them: Have you ever cheated on an exam? Have you ever called in sick when you could have worked? For cases like these, techniques that add an element of chance to the answer have existed since the 1960s. Nobody can then reliably infer a person's behavior from their individual answer, but the proportion in the group as a whole can still be estimated (Jacob et al. 2019, p. 176).

**Randomized response technique (RRT).** The basic idea goes back to Warner in 1965. Each person gets two questions, a sensitive one and an innocuous one, and a random mechanism decides which one they answer. Only they know which one it was (Jacob et al. 2019, p. 176).

Let's look at an invented example. Before the question, 600 students each toss their own coin, out of sight. If it lands heads, they answer "Have you ever cheated on an exam?", and if it lands tails, "Was your mother born in the first half of the year?". A total of 210 answer yes, or 35%. With a fair coin, we expect about 300 people to get the innocuous question. If 50% of them answer yes (an assumption), we can expect about 150 yes answers from them. For the sensitive question, this gives the estimate (210 minus 150) / 300 = 20%. Which people actually answered which question remains unknown. In general, following Schnell (2019, p. 95):

```math
\begin{aligned} \text{estimated proportion} &= \frac{\text{proportion of yes answers} - (1 - p) \cdot q}{p} \\ p &= \text{probability of getting the sensitive question (here } 0.5\text{)} \\ q &= \text{proportion of yes answers to the innocuous question (here } 0.5\text{)} \\ \frac{0.35 - 0.5 \cdot 0.5}{0.5} &= 0.20 \end{aligned}
```

Two things determine whether the protection holds. First, the answer to the innocuous question must not be inferable from other responses. Jacob et al. use "Were you born in an even-numbered year?" as an example (2019, p. 176). But if you also collect age or year of birth, a yes from someone born in an odd-numbered year reveals a yes to the sensitive question. In this example, the questionnaire doesn't otherwise collect the mother's birth month, which is why that question is used. Second, respondents must generate the randomness themselves, out of sight and without leaving a trace, for example with their own coin. Don't let the questionnaire generate the random assignment. Otherwise, the data will show who received which question, and the protection is lost.

The price shows up in precision. In this example, the 95% confidence interval runs from about 12% to 28%, that is, plus or minus 7.6 percentage points. A direct question yielding the same estimated proportion with the same 600 respondents would have a margin of error of plus or minus 3.2 percentage points. For the same precision, you'd need about 5.7 times as many respondents with RRT here, around 3,400 instead of 600 (my own calculation with a simplified normal approximation). For a real analysis, the usual formulas for confidence intervals don't apply, and special procedures are needed (Jacob et al. 2019, p. 179). Schnell (2019, p. 98) and Jacob et al. (2019, p. 179) also stress that RRT requires larger samples.

**Item count technique (list experiment).** Here the group is split at random. Everyone states only how many statements on a list apply to them, not which ones. One half sees three innocuous statements, the other half the same three plus the sensitive one. The difference between the means estimates the proportion. This rests on the assumption that both groups are the same on the innocuous statements and that the difference is due solely to the sensitive statement (Jacob et al. 2019, pp. 177-178; Schnell 2019, p. 96). In the invented example with 300 people in each group: short list mean 1.62, long list mean 1.80, estimated proportion 18%. With assumed standard deviations of 0.85 and 0.95, the confidence interval runs from about 4% to 32% (my own calculation, equally simplified). This provides only a rough estimate.

**Crosswise model.** The person gets two questions, an innocuous one and a sensitive one, for example: "Was your mother born in January, February or March?" and "Have you ever used cannabis?" They only state whether their two answers are the same or different. Neither answer incriminates them, so there's no reason to protect themselves with a "no". No coin is needed. Online, respondents simply select one of two response options. For the estimate, you need to know how many people answer yes to the innocuous question, and this proportion must not be 50% (Jacob et al. 2019, pp. 178-179). In an online experiment, the model produced the highest proportions of sensitive behavior. That may mean it measures best, but also that it produces false positives (Höglinger et al. 2016, cited in Jacob et al. 2019, p. 179).

**What practice shows.** In a methods experiment by Coutts and Jann, up to 20.5% said they didn't fully understand the random mechanism of the RRT, fewer than a quarter were confident that it protects their anonymity, and some answered no every time, just to be safe. The item count technique did better: 91.8% said they had understood it, but here too, only 28.6% believed in their anonymity (Coutts and Jann 2011, cited in Jacob et al. 2019, pp. 177-178). An analysis of 68 comparative studies did find more valid results with RRT, but the differences between the studies are considerable and hard to predict (Lensvelt-Mulders et al. 2005, cited in Schnell 2019, p. 98). In addition, most results are based on studies with students or self-selected online samples (Schnell 2019, p. 99).

My conclusion matches that of Jacob et al.: no single approach is clearly superior, analysis often requires special software, and such methods should be used only when a sensitive question can't be asked using conventional methods (2019, p. 179). For a thesis, a list experiment can still be an interesting addition, as a comparison with the direct question, not as a replacement.

## A close relative: acquiescence bias

Not every overly positive answer is socially desirable. Acquiescence means that respondents agree with statements in a stereotyped way, regardless of their content and even when the statements contradict each other (Möhring and Schlütz 2019, p. 57). Social desirability depends on the content of the question. Acquiescence does not, so it calls for different remedies. Ask for ratings rather than for agreement with statements (Roberts 2016, cited in Möhring and Schlütz 2019, p. 57). Statements worded in both directions can reveal a tendency to say yes, but they bring their own problems with double negatives (ibid.). How the position of questions and answers shifts results as well is covered in the article [Order effects in surveys](/blog/order-effects-in-surveys/).

## Employee survey example: honest ratings of your own manager

Nowhere is social desirability as tangible as in questions about your own manager. If you ask employees naively, you may not be measuring their needs, but what they'd like their boss to believe (Linke 2018, p. 8).

The case study in Wiertalla is revealing. The employees trusted the data collection by the external provider. Their concern lay elsewhere: that managers would work backward from the percentages in their report to figure out who had answered what. Reports there were only produced for managers with at least five responding employees (Wiertalla 2013, pp. 186-187).

> **The danger lies less in the technology than in small groups and in how the results are handled.**

One point is directly relevant to social desirability bias: a high response rate achieved through pressure comes at the price of embellished answers. More on this in the article [Survey response rate](/blog/survey-response-rate/). Minimum group size for reports, excluding small groups and agreeing on all of this with the works council are covered in the article [Employee survey](/blog/employee-survey/#anonymity).

## Sensitive questions in QUESTIONSTAR

A word on our own product, with specific features. No tool eliminates social desirability bias. QUESTIONSTAR helps you set up the survey so that your assurance is true.

**Invitations and reminders.** In the [Access](/help/survey-settings/access/) settings, you decide whether the survey is public, password-protected or available by invitation only, and whether the same person can take part more than once. Personal links contain a randomly generated, pseudonymized ID that doesn't itself reveal personal information ([Recognizing returning respondents](/help/survey-settings/recognizing-returning-respondents/)). You send reminders only to those who haven't responded yet, also in anonymous surveys. Whether QUESTIONSTAR saves the IP address is set in the privacy settings: not at all, in full or partially, in which case the first two blocks are masked and only the last two are visible. A stricter setting can't be reverted later, so nobody gains more access afterward than intended. For an anonymous employee survey, choose "Do not save respondents' IP addresses": on a company network, it's precisely the last two blocks that often identify the individual computer.

**Two levels of anonymity.** You set both in the [privacy settings](/help/survey-settings/privacy/), and to protect respondents, neither can be undone:

| Setting | What happens | Suitable for |
|------|------|------|
| "Save Private Data Separately" | Data is exported in two files with records in different orders. The files can be linked only through an ID. | Confidential: a link remains possible, for example for prize draws or follow-up rounds |
| additionally "Anonymize IDs" | No ID in the export, random record order, reports and analyses only once ten questionnaires have been fully completed | Anonymous: answers can no longer be linked to a person through the technology. Participation status remains known for reminders. |

You also decide which users of your account have access to the results or to the personal details at all.

**Private questions.** You mark questions that collect personal data as "Private Data". Their answers are only shown in the analysis and in the raw data export once at least ten questionnaires have been fully completed (see [privacy settings](/help/survey-settings/privacy/)). If QUESTIONSTAR is connected to an AI assistant via the [MCP interface](/help/integrations/mcp-ai/), questions marked this way aren't transferred to the assistant when the setting is on. Unmarked free text can still contain personal information.

**What QUESTIONSTAR doesn't do.** There's no dedicated question type for randomized response or item count. The crosswise model uses a standard single-choice question with two response options, and you calculate the estimate yourself. You build a list experiment with an [A/B test](/help/branching-and-display-logic/a-b-tests/): the short and the long list are two variants, and each person sees only one. You calculate the difference between the means yourself. You build an RRT question as a yes/no question with instructions for the hidden coin toss, which each person does with their own coin, and again you calculate the estimate yourself. The questionnaire must not generate the random assignment for RRT. Otherwise, the data will show who received which question. No setting can prevent inferences from answer content, free text or very small groups.

## Checklist: preparing sensitive questions

1. For each sensitive question, clarify: privacy, fear of disclosure or social norm? Choose the remedy accordingly.
2. Let people fill in the survey themselves, without an interviewer and ideally without onlookers.
3. Decide: anonymous or confidential? Write exactly what's true in the invitation, including about the participation status.
4. Explain the purpose of sensitive questions in one sentence, offer ranges instead of asking for exact figures, specify a time period and use neutral wording for both answers.
5. Use normalizing sentences only if they're true. Don't use the third person as a substitute for the direct question.
6. Offer "prefer not to answer", and "can't judge" only where knowledge may be lacking.
7. Put sensitive questions toward the end, from the general to the specific.
8. In the pretest, ask which answers feel "desirable".
9. Set the minimum group size for reports and the characteristics you'll ask for in advance, and keep to them yourself.
10. Use indirect techniques only for individual, unavoidably sensitive questions and with enough respondents. With RRT, respondents generate the randomness themselves, out of sight and without any record.

## Frequently asked questions about social desirability bias

### What is social desirability bias?

Social desirability bias is the tendency of respondents to shape their answers around what's considered acceptable in their environment, which distorts the true value (Möhring and Schlütz 2019, p. 60). Desirable behavior tends to be exaggerated. Undesirable behavior tends to be played down.

### What is an example of social desirability bias?

Voter turnout and donations tend to be overreported in surveys. Alcohol consumption and illness tend to be underreported (Bradburn et al. 2004, cited in Schnell 2019, pp. 44-45). Questions about income are sensitive too. The roughly 55% refusals in the 1990 ZUMA Bus survey show this sensitivity, but on their own they don't prove socially desirable false answers (Jacob et al. 2019, p. 175).

### How can you reduce social desirability bias?

You can't avoid it completely, but you can reduce it (Möhring and Schlütz 2019, p. 64). Important starting points are a self-administered survey, an assurance of anonymity that's true, and neutral questions. Indirect techniques such as randomized response are only worth it for individual, highly sensitive questions.

### How do you ask sensitive questions in a survey?

It's best to have respondents fill in sensitive questions themselves, without an interviewer, and to promise only the anonymity that's actually true (Möhring and Schlütz 2019, p. 62). Offer ranges instead of exact figures, word questions so that no answer seems embarrassing, and put sensitive questions toward the end (Jacob et al. 2019, p. 175; Schnell 2019, p. 109). Offer "prefer not to answer" and record it separately from "don't know" (Schnell 2019, p. 46).

### Are online surveys less prone to social desirability bias?

As a rule, yes. In self-administered surveys, by mail or online, the effect is smaller than in telephone or face-to-face interviews (Möhring and Schlütz 2019, p. 62). The condition is that respondents believe the assurance of anonymity or confidentiality.

### What is the difference between anonymous and confidential surveys?

Anonymous means the answers can't be linked to any person. Confidential means a link would be possible, but it isn't made and the data is protected. If you work with personal links or IDs, you have to say exactly which case applies (Schnell 2019, p. 296). With personal links, it remains known who has already taken part, even when the answers are anonymous.

### What is the randomized response technique?

In the randomized response technique, a random mechanism, such as a coin toss, decides whether someone answers the sensitive question or an innocuous one. The proportion of the sensitive behavior in the group can be estimated. An individual answer, however, doesn't allow a clear conclusion about whether the person has shown this behavior (Jacob et al. 2019, pp. 176-177). The technique requires considerably more respondents (ibid., p. 179). The crosswise model works without a coin and runs online as an ordinary question (ibid., pp. 178-179).

## Sources

Bradburn, N. M., Sudman, S. and Wansink, B. (2004). Asking Questions. The Definitive Guide to Questionnaire Design (Revised edition). San Francisco: Jossey-Bass. Cited in Schnell (2019), pp. 44-45.

Coutts, E. and Jann, B. (2011). Sensitive Questions in Online Surveys: Experimental Results for the Randomized Response Technique (RRT) and the Unmatched Count Technique (UCT). Sociological Methods & Research, 40, 169-193. Cited in Jacob et al. (2019), pp. 177-178.

Höglinger, M., Jann, B. and Diekmann, A. (2016). Sensitive Questions in Online Surveys: An Experimental Evaluation of Different Implementations of the Randomized Response Technique and the Crosswise Model. Survey Research Methods, 10(3), 171-187. Cited in Jacob et al. (2019), p. 179.

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.

Krumpal, I. (2013). Determinants of social desirability bias in sensitive surveys: a literature review. Quality & Quantity, 47(4), 2025-2047. Cited in Möhring and Schlütz (2019), pp. 60 and 62.

Lensvelt-Mulders, G. J. L. M., Hox, J. J., van der Heijden, P. G. M. and Maas, C. J. M. (2005). Meta-Analysis of Randomized Response Research. Thirty-Five Years of Validation. Sociological Methods & Research, 33(3), 319-348. Cited in Schnell (2019), p. 98.

Linke, R. (2018). Mitarbeiterbefragungen optimieren. Von der Befragung zum wirksamen Management-Instrument [Optimizing employee surveys: from survey to effective management tool]. Wiesbaden: Springer Gabler. In German.

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, slides 106, 108 and 110 (chapter 3). QUESTIONSTAR. [Course, chapter 3](/principles-of-survey-research/chapter-3/slide-108.html).

Möhring, W. and Schlütz, D. (2019). Die Befragung in der Medien- und Kommunikationswissenschaft. Eine praxisorientierte Einführung [Surveys in media and communication studies: a practical introduction] (3rd ed.). Wiesbaden: Springer VS. In German.

Roberts, C. (2016). Response styles in surveys: Understanding their causes and mitigating their impact on data quality. In C. Wolf, D. Joye, T. W. Smith and Y. Fu (eds.), The SAGE Handbook of Survey Methodology (pp. 579-596). Thousand Oaks: Sage. Cited in Möhring and Schlütz (2019), p. 57.

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.

Tourangeau, R., Rips, L. J. and Rasinski, K. (2000). The Psychology of Survey Response. Cambridge: Cambridge University Press. Cited in Schnell (2019), p. 44, n. 39.

Warner, S. L. (1965). Randomized Response: A Survey Technique for Eliminating Evasive Answer Bias. Journal of the American Statistical Association, 60, 63-69. Cited in Jacob et al. (2019), p. 176.

Weinreich, U. and von Lindern, E. (2008). Praxisbuch Kundenbefragungen [Customer surveys: a practical handbook]. München: mi-Fachverlag, FinanzBuch Verlag. In German.

Wiertalla, K. (2013). Der Einsatz von Mitarbeiterbefragungen als Lerninstrument in Unternehmen. Dargestellt anhand des Praxisbeispiels der Mitarbeiterbefragung bei der Ingram Micro Distribution GmbH [Employee surveys as a learning tool in companies, illustrated by the employee survey at Ingram Micro Distribution GmbH]. Dissertation, University of Siegen. In German.

Yeager, D. S. and Krosnick, J. A. (2012). Does Mentioning "Some People" and "Other People" in an Opinion Question Improve Measurement Quality? Public Opinion Quarterly, 76(1), 131-141. Cited in Möhring and Schlütz (2019), p. 63.

**Ready for your next survey?**

Survey people anonymously, invite them personally and remind only those who haven't responded yet. Start for free, no credit card required.

[Start for free](/order/?code=free)

I hope your respondents feel comfortable telling you even the uncomfortable truths, and that your questionnaire makes it easy for them.

**Related articles:**

- [Employee survey: process, questions and analysis](/blog/employee-survey/)
- [Order effects in surveys: what they distort and how to avoid them](/blog/order-effects-in-surveys/)
- [Data protection in online surveys](/blog/data-protection-in-online-surveys/)

*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.

## Related articles

- [employee-survey](/blog/employee-survey)
- [order-effects-in-surveys](/blog/order-effects-in-surveys)
- [data-protection-in-online-surveys](/blog/data-protection-in-online-surveys)

*Author: Dr. Paul Marx — see [About](/about).*