Research Bias: Types, Examples, and How to Avoid It

Every research study is vulnerable to bias. Research bias is any systematic distortion in how a study is designed, conducted, analyzed, or reported that pulls the results away from the truth. Bias isn't the same as random error, and adding participants won't fix it. It has to be designed out from the start, or at least named honestly in the limitations section. Reviewers know what to look for, and a methodology section that doesn't address bias is one of the most common reasons manuscripts get sent back.

This guide explains the five main types of research bias, with examples of each and the prevention methods that work in practice. Each section shows where bias enters a study, so you can check your own design before you submit. For the wider framework, see our research methodology guide.

What Are the Main Types of Research Bias?

There are five main types of research bias, grouped by where in the research process the distortion enters. Understanding the categories helps you audit your own study and recognize which biases are most likely to threaten your specific design. Different research approaches also face different bias profiles. Our companion guide on quantitative vs qualitative research covers how the two differ in what they're most vulnerable to.

Selection and Sampling Bias

Selection bias happens when the people in your study don't represent the population you want to draw conclusions about. It's one of the most common threats to validity in social science, because most studies use convenience samples rather than random ones. The result can be a study whose internal logic holds but whose broader conclusions don't. The relationship between population and sample shapes what you can claim, and our companion guide on population vs sample in research covers it in full.

Closely related is sampling bias, which occurs when the sampling procedure itself systematically excludes or under-represents certain groups. A survey distributed only through email can't capture households without internet access. A study recruiting at a single university campus will under-represent older adults. Each design choice excludes someone, and the methodology section needs to name who.

• Self-selection bias. People who volunteer for a study often differ from those who don't. A workplace satisfaction survey that relies on volunteers may draw mostly employees with strong opinions, good or bad.
• Nonresponse bias. When a meaningful portion of selected participants doesn't respond, the responders may differ from non-responders in ways that affect the results. A low response rate raises the risk, but no response rate guarantees unbiased results. What matters is whether nonrespondents differ from respondents on the outcome. Our guide to nonresponse bias covers how to check.
• Survivorship bias. Studying only the cases that "survived" some process produces distorted conclusions. Studying successful entrepreneurs to identify success factors ignores the much larger group of failed entrepreneurs who may have had the same characteristics.
• Healthy worker effect. In occupational health research, workers often look healthier than the general public. That's because people who are seriously ill are less likely to be working.

The main defense is probability sampling, where every member of the population has a known chance of being chosen. Random assignment and targeted follow-up of the groups least likely to respond also help. Where you can't remove selection bias, describe how you sampled and say who your results do and don't apply to.

Our detailed guide on selection bias covers this category in full. It includes the subtype taxonomy, attrition bias in longitudinal designs, worked detection steps, and how to write selection bias into a limitations section.

Information and Measurement Bias

Information bias comes from how data is collected, measured, or recorded. Even a perfectly representative sample can give wrong answers if the measurements are off in a consistent way. Strong studies use tested instruments, trained data collectors, and blinded measurement to keep it in check.

• Recall bias. People asked to remember past behaviors or exposures often misremember in a consistent pattern. Those who've had a bad outcome, such as an illness, may recall more past risk factors than people who haven't.
• Observer bias. Researchers who measure outcomes can read unclear data in line with what they expect, without realizing it. Blinded measurement (where the data collector doesn't know which condition the participant is in) is the standard prevention.
• Measurement error. Instruments can be off in a consistent direction (systematic error) or vary at random around the true value (random error). Systematic error is the bias problem. Random error adds noise but doesn't shift the answer.
• Instrument bias. A survey scale built for one group may give misleading results when used with another. A financial literacy scale validated on US adults may not work for adolescents or for adults in different financial systems.

To prevent it, use instruments that have been tested and shown to work for your population. Train everyone who collects data so they follow the same steps. Blind measurement where you can, and check key measures with more than one method. Studies in a second language or another culture need extra care when instruments are translated.

Our detailed guide to information bias covers this category in full, with a scenario table for identifying each type. Separate guides cover recall bias, observer bias, and measurement error.

Response Bias

Response bias occurs when participants answer questions in ways that don't reflect their true beliefs, behaviors, or experiences. Self-report data is particularly vulnerable. Even with valid instruments and unbiased samples, what participants say can systematically differ from what's actually true.

• Social desirability bias. People give answers they think will look good rather than answers that match what they really think or do. It's a major problem in research on drug use, sexual behavior, racial attitudes, and income.
• Acquiescence bias. Some people tend to agree with whatever statement they see, no matter what it says. Mixing positively and negatively worded items in a scale helps expose it.
• Demand characteristics. Participants pick up cues about what the researcher expects to find and adjust their behavior to fit. A related but separate effect is the Hawthorne effect, where people change their behavior because they know they are being observed. The name comes from studies at Western Electric's Hawthorne Works in the 1920s and 1930s. Later reanalyses of the original data question how large that effect really was (Levitt and List, 2011).
• Extreme response style and central tendency bias. Some respondents prefer to use the extremes of any rating scale; others avoid the extremes and cluster in the middle. Both patterns can distort group comparisons if they're concentrated in one group more than another.

The best defense is true anonymity, which sharply cuts social desirability bias on sensitive topics. Neutral wording and reverse-scored items help too, since they expose people who agree with everything. Where you can, check what people say against what they do.

Note that response bias and nonresponse bias are different threats. Response bias distorts the answers you received. Nonresponse bias distorts who is in your dataset at all. Our detailed guide on response bias covers the category in full. It includes detection methods, survey mode effects, and the anonymity distinction that drives most of the prevention. Our guide to nonresponse bias covers how to detect and report the selection side.

Researcher Bias

Researcher bias refers to systematic distortions introduced by the researcher's own expectations and decisions during data collection, analysis, and reporting. Unlike the categories above, this one isn't about participants. It's about how the person running the study shapes what it finds.

• Confirmation bias. Researchers tend to find evidence that confirms their hypotheses and overlook or discount evidence that contradicts them. This isn't usually intentional. It reflects how human cognition works under conditions of complex data and time pressure.
• Observer expectancy effect. Researchers who know which participants are in the experimental versus control condition can unconsciously treat them differently or interpret their responses differently. Double-blinding (where neither the participant nor the data collector knows the condition) is the gold standard prevention. Our guide to observer bias covers blinding in detail.
• P-hacking and data dredging. Running many statistical tests until something reaches significance, or trying multiple analytic approaches and reporting only the one that produced the desired result. Pre-registration is the strongest prevention.
• Researcher allegiance and funding bias. Researchers strongly committed to a theory or therapy tend to find results that favor it. So do studies paid for by groups with a financial stake in the outcome. Disclosure is required, and blinding and pre-registration add protection.

The main defense is to pre-register your hypotheses and analysis plan before you collect data. Blind the analysis where you can, support replication, and disclose all funding and conflicts of interest.

Publication Bias

Publication bias works at the level of the literature, not within a single study. Studies with significant or striking results are more likely to be submitted and accepted than studies with null results. As a result, the published record shows more positive findings than really exist. That skews meta-analyses, systematic reviews, and what readers take to be the evidence.

• The file drawer problem. Studies with null results often end up in researchers' file drawers rather than journals. Robert Rosenthal named the problem in 1979. The published literature shows what worked; the unpublished literature shows what didn't, and readers see only the former.
• Outcome reporting bias. Within a single study, researchers may choose to report only the outcomes that produced significant results, leaving non-significant outcomes out of the manuscript. Pre-registration of all planned outcomes counteracts this.
• Time-lag bias. Studies with positive results are published faster than studies with null results. At any given moment, the recent literature is even more skewed toward positive findings than the cumulative literature.
• Language bias. Studies published in English are more accessible to international meta-analyses than studies published in other languages, even when the research is equally rigorous. Systematic reviews need to address this in their search strategies.

At the system level, prevention needs journals willing to publish null results and registries such as the Open Science Framework. Registered reports help too, because the journal agrees to publish based on the methods before results are known. In your own study, pre-register every outcome and report them all, significant or not.

Our detailed guide on publication bias covers this category in full. It includes funnel plots and other detection methods, registered reports, and how to handle publication bias in a systematic review.

Research Bias Examples: Which Type Is It?

Research bias examples are easiest to learn by matching a scenario to its type. Each row below names the bias and the category it belongs to.

How to Audit Your Study for Bias

Before submitting your manuscript, work through this audit. Each step corresponds to a stage where bias commonly enters and is most easily addressed before peer review begins.

1. Audit your sampling. Define your target population. Describe the sampling procedure that produced your sample. Calculate your response rate. Identify systematic differences between responders and non-responders where possible. Document who the results can and can't generalize to.
2. Audit your measurement. List every instrument and confirm each was tested for your population. Describe how data collectors were trained. Say whether measurement was blinded, and note any known error in your instruments.
3. Audit your participants' answers. For self-report data, list the items most likely to draw socially desirable answers. Confirm how you protected anonymity. Note whether you used reverse-scored items or attention checks.
4. Audit your own decisions. Were your hypotheses pre-registered? Did you decide your analytic approach before seeing the data? Did you run additional analyses after the initial results, and if so, did you report them all?
5. Audit your reporting. Does your manuscript report all pre-registered outcomes, including non-significant ones? Have you disclosed all funding sources and potential conflicts of interest?
6. Address bias in your limitations. Where bias couldn't be eliminated, name it explicitly. Reviewers are more confident in studies that acknowledge bias transparently than in studies that pretend none exists.

Common Mistakes Researchers Make About Bias

The same misunderstandings about bias appear repeatedly in graduate research. Knowing them in advance saves a round of revisions.

• Confusing bias with random error. Random error increases noise; bias shifts the answer. Bigger samples reduce random error but don't fix bias. A biased study with 10,000 participants is still biased.
• Treating "no significant difference" as evidence of no bias. Failing to find an effect of bias doesn't mean bias is absent. It often means the study wasn't powered to detect it.
• Pretending limitations don't exist. Reviewers know every study has limits. Papers that name them honestly do better than papers that don't.
• Confusing internal validity with external validity. A randomized experiment can show cause and effect within the study (internal validity). It may still not apply to other people or settings (external validity). Each needs its own defense.
• Treating the response rate as a bias test. A high response rate lowers the risk of nonresponse bias but doesn't rule it out. What matters is whether the people you missed differ on your outcome.
• Assuming peer review catches bias. Peer reviewers are time-pressed and unfamiliar with your specific data. The primary responsibility for identifying and addressing bias rests with the researcher, not the reviewer.

Professional Editing for Your Research Manuscript

Reviewers screen the methods and limitations sections for bias before they read the statistics. A clear bias discussion shows what you controlled, what you couldn't, and how much weight your conclusions can bear.

Editor World provides journal article editing, dissertation editing, and academic editing. Every editor is a native English speaker from the United States, the United Kingdom, or Canada. Every document is edited by a person, never by AI. Choose an editor with experience in your field. The first 300 words of any document are edited at no cost, no commitment. A certificate of editing confirming human-only editing is available as an optional add-on.

Frequently Asked Questions About Research Bias

Related Guides

Our guide to selection bias covers sampling, self-selection, attrition, and survivorship bias with worked examples.

The guide to nonresponse bias explains why response rates can mislead and how to run a wave analysis.

For distortion in the answers people give, read our guide to response bias, including social desirability and acquiescence.

Our information bias guide covers measurement problems, with companion guides on recall bias, observer bias, and measurement error.

The guide to publication bias explains funnel plots, registered reports, and the file drawer problem.

The research methodology guide shows how bias prevention fits into the overall design of a study.

Reviewed by the Editor World editorial team under the direction of Patti Fisher, PhD, founder. Editor World has provided 100% human editing by native English editors since 2010. About Editor World