Types of bias in research
Bias in research is a systematic error that pushes results away from the truth in a consistent direction. The most common types are selection, sampling, non-response, attrition, recall, measurement (information), observer, social desirability, confounding and publication bias. Unlike random error, bias does not shrink as your sample grows, so you have to prevent it through design rather than fix it with more data.
Get the free Research Design Decision Tree
The short answer
Imagine weighing everyone in a study on a scale that reads two kilograms light. Weigh a thousand people instead of a hundred and the average will still be two kilograms wrong. That is bias: an error with a direction. It can enter when you choose participants, when you measure them, when you lose them during follow-up, when you analyze the data, and even when the results are published.
Long lists of biases exist. Sackett (1979) catalogued dozens of them, grouped by the stage of research where they occur, and many later lists build on his. You do not need to memorize them all. You need to recognize the common ones, know which threaten your own design, and explain in your methods and limitations what you did about them.
Bias vs random error
| Bias (systematic error) | Random error (chance) | |
|---|---|---|
| What it does | Shifts results consistently in one direction | Scatters results around the true value |
| Effect of a larger sample | None: the error stays | Shrinks as the sample grows |
| What it threatens | Validity (accuracy) | Precision (reliability) |
| How you handle it | Design: sampling, blinding, randomization, standard procedures | Larger samples, repeated measures, confidence intervals |
| Shown by statistics? | No: a narrow confidence interval can be precisely wrong | Yes: reflected in standard errors and confidence intervals |
This is the same distinction that separates validity from reliability; see the validity and reliability guide.
Common types of bias at a glance
| Bias | What it is | Example | How to reduce it |
|---|---|---|---|
| Selection bias | The people who enter the study, or a comparison group, differ systematically from those they are meant to represent or be compared with. | Patients well enough to attend a clinic are compared with all patients in the community. | Clear eligibility criteria, consecutive or random recruitment, random allocation with concealment. |
| Sampling bias | The sampling method makes some members of the population more likely to be included. | An online survey on digital skills recruits only through social media. | Probability sampling from a full sampling frame; describe who is missing. |
| Non-response bias | People who decline to take part differ from those who respond. | Staff with the heaviest workloads do not return a burnout questionnaire. | Reminders, short instruments, several contact modes; compare responders with non-responders on known characteristics. |
| Attrition bias | Participants who drop out differ from those who stay, or dropout differs between groups. | Those not improving leave a therapy trial more often than those who are. | Retention planning, recording reasons for dropout, intention-to-treat analysis, appropriate handling of missing data. |
| Recall bias | Accuracy of remembered information differs between groups. | Parents of children with a condition search their memory harder for past exposures than parents of healthy children. | Records instead of memory, prospective data collection, structured questions with fixed time frames. |
| Measurement (information) bias | The way an exposure or outcome is measured is systematically wrong, or differs between groups. | Blood pressure taken with different devices in the two arms of a study. | Validated instruments, calibration, the same procedure for every group, trained assessors. |
| Observer bias | The assessor’s expectations influence what they record. | A researcher who knows which pupils received an intervention rates their behavior more favorably. | Blind assessors to group, use objective outcomes, check inter-rater agreement. |
| Social desirability bias | Participants answer in the way they think is acceptable rather than what is true. | Respondents under-report alcohol use in a face-to-face interview. | Anonymous or self-administered formats, neutral wording, assurance of confidentiality. |
| Confounding | A third factor linked to both exposure and outcome distorts the apparent relationship. | Coffee appears linked to heart disease because coffee drinkers in the sample smoke more. | Randomization, restriction, matching, measuring known confounders and adjusting in analysis. |
| Publication bias | Studies with positive or striking results are more likely to be published. | A review finds only trials favoring a treatment because null trials were never published. | Search trial registries and grey literature; register your own protocol and report all outcomes. |
Terms overlap. Recall, observer and social desirability bias are all forms of information bias, and sampling and non-response bias are forms of selection bias. What matters in a thesis is that you name the specific mechanism that threatens your study.
Bias by stage of a study
- Before data collection: sampling bias, selection bias, non-response bias. The sampling methods guide covers how sampling choices create or limit them.
- During data collection: measurement, observer, recall and social desirability bias.
- During follow-up: attrition bias.
- During analysis: confounding that is not addressed, and selective reporting of the outcomes or analyses that “worked”.
- After the study: publication bias, which matters most if you are doing a systematic review.
How to reduce bias, step by step
- List the biases your design is prone to. A case-control study invites recall bias; a cohort study invites attrition bias; a survey invites non-response and social desirability bias. Write the list before you finalize your methods.
- Fix the sample. Define the target population, use a sampling frame that covers it, and set eligibility criteria in advance.
- Randomize and conceal allocation if you are comparing interventions. Randomization balances known and unknown confounders on average; concealment stops recruiters from steering participants into groups.
- Blind where you can. Participants, those delivering the intervention and outcome assessors. If full blinding is impossible, blind the assessor.
- Standardize measurement. Validated instruments, a written protocol and training so that every participant is measured the same way.
- Plan for missing data. Build retention into the study, record why people leave, and state in advance how you will analyze incomplete data.
- Pre-specify outcomes and analyses. A registered protocol or analysis plan guards against selective reporting.
- Report what remains. No study is free of bias. Say which biases may still affect your results and in which direction.
If you are appraising other people’s studies, use a structured tool. For randomized trials the standard is the Cochrane RoB 2 tool (Sterne et al., 2019), which covers bias from the randomization process, deviations from intended interventions, missing outcome data, measurement of the outcome and selection of the reported result. The risk of bias assessment guide explains how to use it and which tools fit other designs.
Worked example: a school exercise study
Suppose you are a public health master’s student planning a study of whether a daily ten-minute classroom exercise break improves pupils’ concentration over one school term. Several schools in one district have agreed to take part. This example covers planning only; there are no results.
- Selection bias. Letting head teachers volunteer for the intervention would put enthusiastic schools in one group. You randomize schools to intervention or control, and someone outside the team holds the allocation list.
- Confounding. Schools differ in size and socioeconomic intake. You record these characteristics and plan to adjust for them, and you note that with few schools randomization may not balance them well.
- Observer bias. Teachers know whether their class had the breaks, so their ratings of concentration could follow expectations. You add a short computer-based attention task scored automatically, and you train research assistants who do not know school allocation to supervise it.
- Non-response bias. Parental consent may come mostly from engaged families. You send consent forms in the main languages spoken locally and compare consenting and non-consenting classes on available school-level data.
- Attrition bias. Pupils who move schools or are often absent may differ from others. You record reasons for missing data and plan an intention-to-treat analysis.
- Selective reporting. You name the primary outcome and time point in a protocol before data collection.
Notice that each bias is paired with a specific design step. That pairing is what examiners look for in a methods chapter.
Weak vs strong limitations statements
| Weak | Why it fails | Stronger |
|---|---|---|
| The study may have some bias. | Names no mechanism and no direction. | Because outcome assessors knew group allocation, observer bias may have inflated ratings in the intervention group. |
| The large sample removed bias. | Sample size reduces random error, not bias. | The large sample gives precise estimates, but recruitment through one online channel may have excluded people with limited internet access. |
| Participants answered honestly. | An assumption, not a safeguard. | The questionnaire was anonymous and self-administered to reduce social desirability bias, though some under-reporting remains possible. |
| Results were adjusted for confounders. | Does not say which ones, or what was not measured. | Analyses adjusted for age, sex and school size; physical activity outside school was not measured and may confound the association. |
Common mistakes
- Confusing bias with random error. A significant p-value or narrow confidence interval says nothing about bias.
- Listing biases without acting on them. A generic list in the limitations section, with no design steps in the methods, reads as an afterthought.
- Ignoring direction. Say whether a bias would likely make your effect look bigger or smaller, where you can reason about it.
- Assuming randomization fixes everything. It addresses confounding at allocation, not measurement bias, attrition or selective reporting.
- Forgetting non-response. A low response rate is not itself bias, but it opens the door to it. Compare responders with non-responders where you can.
- Calling every limitation “bias”. A small sample or a short follow-up is a limitation, but not necessarily a source of systematic error.
Checklist
- You have listed the biases your design is most prone to.
- Each is matched to a specific step in your methods.
- Your sampling frame and eligibility criteria are defined in advance.
- Allocation is randomized and concealed if you compare interventions.
- Assessors are blinded where possible, and instruments are validated.
- You have a plan for dropout and missing data.
- Primary outcomes and analyses are pre-specified.
- Your limitations name remaining biases and their likely direction.
Get the free Research Design Decision Tree
A step-by-step decision tree from your research question to the design that fits it — qualitative, quantitative, or mixed methods — with the sampling approach each path implies, five matched question-to-design examples, and a ten-point pre-commitment checklist, from Research Design Simplified. We’ll email you the download link.
Frequently asked questions
What are the main types of bias in research?
Common types are selection bias, sampling bias, non-response bias, attrition bias, recall bias, measurement or information bias, observer bias, social desirability bias, confounding and publication bias. They arise at different stages, from recruiting participants to publishing results.
What is the difference between bias and random error?
Bias is systematic error: it pushes results in one direction and does not shrink with a larger sample. Random error is chance variation that scatters results around the true value and becomes smaller as sample size increases. Bias threatens validity; random error threatens precision.
Is confounding a type of bias?
Many epidemiology texts treat confounding as a type of bias, because a third factor linked to both exposure and outcome distorts the apparent effect. Others list it separately from selection and information bias. Either way, you deal with it through design (randomization, matching, restriction) or analysis (adjustment).
How can you reduce bias in research?
Choose a design that limits bias for your question, define and recruit your sample carefully, use validated measures, blind assessors and participants where possible, standardize procedures, maximize follow-up, register a protocol and analysis plan, and report all outcomes.
What is the difference between selection bias and sampling bias?
Sampling bias is a form of selection bias in which the way you sample makes some members of the population more likely to be included, so the sample does not represent the population. Selection bias is broader and also covers how participants enter or leave comparison groups.
What tool is used to assess risk of bias in randomized trials?
The Cochrane risk-of-bias tool for randomized trials, RoB 2 (Sterne et al., 2019), assesses bias arising from randomization, deviations from intended interventions, missing outcome data, outcome measurement and selection of the reported result.
Who first catalogued types of bias in research?
An early and widely cited catalogue is Sackett’s 1979 paper on bias in analytic research, which listed dozens of biases by the stage of research where they arise. Many later lists build on it.
The Systematic Review and Meta-Analysis Protocol Workbook
Systematic reviews and meta-analyses sit at or near the top of most evidence hierarchies. But while Cochrane provides the methodology, they don't provide the blank templates you need to actually document every step for publication.
60 fill-in templates. Paperback $31.99 on Amazon
See inside the workbookJust want the templates? The Systematic Review & Meta-Analysis Protocol Kit: every template as a fillable PDF, $19. Or all 17 kits for $99.