Types of Bias in Research and How to Reduce Each One

The main types of bias in research are social desirability bias, interviewer bias, leading questions, sampling bias, recall bias and confirmation bias. Each one pulls your findings away from what participants think or do, and each has a practical fix you can build into the sample, the discussion guide or the analysis.

No study is free of bias. The goal is to know which biases your design is exposed to, reduce what you can, and read the results with the rest in mind. Below are the ones that do the most damage in qualitative and mixed-method research.

Types of bias in research at a glance

Social desirability bias

Social desirability bias is the pull to give the answer that makes you look reasonable, generous or healthy. It is strongest on money, politics, relationships, health and anything people feel judged about, which is often the subject of the research.

The presence of another person drives much of it. In our AI interviewer vs human moderator study, 34% of participants said they are more honest with an AI interviewer, 6% said they are more honest with a human, and 60% said it's about the same. 46% said they are more honest with AI specifically because it takes away judgment, and 28% described holding back with humans to manage how they came across. As one participant put it: "AI isn't going to judge me [...] with a human I have to watch my words."

Topic choice followed the same line. 58% would choose an AI interviewer for political or religious topics and 46% for personal finances. Health was the exception, where a human moderator did best at 32%, because participants wanted someone who could recognize distress and refer them to help.

To reduce it, phrase questions so every answer sounds normal, let people answer on their own time, and use AI moderation for sensitive topics while keeping a human option for emotionally heavy ones.

Interviewer bias and moderator effects

Interviewer bias comes from the moderator. A nod at one answer, a pause after another, a probe asked of some participants and skipped for others, or a hypothesis the moderator wants confirmed can all shape what people say.

In the AI interviewer study, comfort with a human moderator came from the person: 68% credited the moderator's warmth, empathy and rapport. Comfort with the AI interviewer came from the absence of a person, with 60% saying it felt less judgmental and safer. One participant said: "With an AI, you can program it to be neutral." The trade-off is warmth, with 26% finding AI more impersonal and 22% feeling it lacked empathy or nuance.

Whoever moderates, the fix is a guide that specifies neutral probes and applies them to everyone. Our guide to writing a qualitative interview guide covers how to write probes that don't steer.

Leading questions and measurement bias

A leading question suggests its own answer, as in "How much did you enjoy the new design?" Order can lead too. Show your message before asking how people choose, and you will hear it played back.

Prompted and unprompted questions give different answers to the same topic. In our apparel brand positioning study, 94% rated comfort and fit as extremely or very important, but only 38% mentioned comfort unprompted when describing how they choose a brand. A study that only asked the rating question would overstate how much comfort drives choice.

Scales carry their own bias. In our brand tracking study, a tracker reported 98% brand love for eight consecutive quarters while the brand lost market share. Nobody had written a leading question. A 5-point scale "stays artificially high and captures a feeling closer to general positive regard," and in a separate study of 615 customers, the 5-point scale inflated brand love by over 5x. Asking people to describe the brand in their own words, then coding the answers, avoids much of that inflation.

Sampling and selection bias

Sampling bias happens when the people you hear from differ in some systematic way from the people you need to understand. Loose screeners and convenience samples are the usual causes.

Exclusion should be deliberate and reported. In our Super Bowl ad testing study, viewers were screened out if they spent the breaks getting food, muted the ads or didn't remember paying attention to them. That made the sample right for testing ad recall, and it means the results describe attentive viewers, not everyone who watched.

Report drop-off within a study too. In the same study, 56 viewers answered the recall questions and 51 the celebrity and AI questions, and each percentage used its own base.

Recall bias

People forget, compress timelines and rewrite past decisions to match what they believe now. Recall bias grows with the time since the event and shrinks when you ask about a specific instance instead of a general habit.

The Super Bowl study ran within 24 to 48 hours of the game for this reason. For purchase research, ask participants to walk through their most recent purchase step by step, checking order history as they talk.

Confirmation bias and bias in qualitative research analysis

Confirmation bias affects the people reading the data. Analysts notice quotes that fit the hypothesis, and stakeholders discount findings that don't. In our study of 76 market researchers, 42% of those who described a time stakeholders pushed back on findings said the findings contradicted what stakeholders already believed.

Analytical shortcuts cause similar problems. In our restaurant message testing study, two messages both made 51.0% of diners more likely to visit, but one pushed 13.5% of diners the other way, almost twice the 7.3% for the other. Reading only the top score would hide that.

To reduce bias in qualitative research analysis, code the answers before forming conclusions, count how often each theme appears, keep every quote linked to its participant, and look for cases that contradict the main story. Pairing each score with an open follow-up also helps, as covered in qualitative vs quantitative research.

On Listen, the AI interviewer asks the same neutral probes of every participant and every theme links back to the participant's own words, which makes it easier to check an interpretation against what people said.

FAQ

What are the main types of bias in research?

Six cause the most trouble in qualitative and mixed-method research: social desirability bias, interviewer bias, leading questions, sampling bias, recall bias and confirmation bias. Each distorts findings at a different stage, from who you recruit to how you read the data.

What is social desirability bias?

Social desirability bias is the tendency for participants to give answers that make them look good rather than answers that reflect what they think or do. It is strongest on sensitive topics such as money, politics and health.

How do you reduce interviewer bias?

Use a written guide with neutral, pre-written probes and apply it the same way in every interview. Avoid reacting to answers, and review transcripts for places where the moderator steered the participant.

What are the types of bias in qualitative research?

Qualitative research is most exposed to interviewer bias, social desirability bias, leading questions, small or skewed samples, recall bias and confirmation bias during analysis. Consistent guides, careful screeners and systematic coding reduce most of them.

Does an AI interviewer reduce bias?

It can reduce social desirability and interviewer bias. In Listen's study of 50 US adults, 34% said they were more honest with an AI interviewer and 6% more honest with a human, mostly because AI felt less judgmental.