How Research Bias Works | Selection, Measurement, Confounding & Reporting Bias

Research bias works by systematically bending the path between the world a study is trying to understand and the estimate, comparison or conclusion the study produces. Unlike ordinary random variation, bias does not necessarily shrink when the sample becomes larger; a huge study can estimate the wrong quantity with extraordinary precision if the evidence pathway is systematically distorted.

Bias is one of those words that becomes less useful when it is used for everything.

In everyday conversation, bias may mean prejudice, preference, political leaning or an unfair tendency.

In research methodology, the word has a narrower job.

It refers to systematic error: a process that tends to move an estimate, comparison or body of evidence away from the target it is supposed to represent.

The distinction matters because random error and systematic error behave differently. If repeated measurements wobble randomly around a target, more information can often increase precision. If the instrument is calibrated five degrees too high, measuring ten thousand times does not make the temperature right.

The governing question: where could the evidence pathway systematically favour one result, one group, one measurement, one analysis or one visible conclusion over the target the study actually claims to represent?

Quick Read

TARGET QUESTION → ELIGIBILITY → SAMPLING → ASSIGNMENT / EXPOSURE → MEASUREMENT → FOLLOW-UP → DATA PROCESSING → ANALYSIS → REPORTING → PUBLICATION → SYNTHESIS

Bias can enter at every arrow.

PREVENT → BLIND WHERE POSSIBLE → RANDOMISE WHERE APPROPRIATE → STANDARDISE → MEASURE VALIDLY → PRESERVE FOLLOW-UP → PRE-SPECIFY → REPORT COMPLETELY → AUDIT → SENSITIVITY TEST → REPLICATE → SYNTHESISE

The goal is not to make a study “bias-free”. That phrase is usually too strong. The practical goal is to identify important routes of systematic distortion, design against them where possible, measure or model their consequences where possible, and stop the final claim from pretending the remaining risk is zero.

1. Bias Is Directional Error, Not Mere Imperfection

Every study is imperfect.

A scale rounds. A respondent forgets. A sample differs by chance. An analyst estimates rather than knows.

Bias becomes the concern when the imperfections are structured in a way that systematically favours particular values, groups or conclusions.

If a weighing instrument adds a random gram in either direction, repeated measurements scatter.

If it always adds five grams, the measurement is systematically shifted.

Research bias is the larger-system version of that shift.

2. Random Error and Bias Need Different Repairs

Problem	Typical effect	Can a larger sample fix it?
Random sampling variation	Estimate fluctuates around the target	Often reduces with more information under the same valid design
Measurement noise	Less precise observations	Sometimes, depending on structure
Systematic selection bias	Observed sample differs from target in outcome-relevant ways	No, not automatically
Systematic measurement bias	Values shifted or classified differently	No, not automatically
Confounding	Exposure-outcome relationship mixed with another causal pathway	No, not automatically

This is why “we have a huge dataset” is not a complete quality argument.

More precision around a biased estimate can make false confidence worse.

3. Bias Is Relative to a Target

A study cannot be called biased in the abstract without asking what it is trying to estimate.

A sample of professional athletes is not “biased” if the target population is professional athletes.

It becomes inappropriate if the conclusion quietly expands to all adults.

A hospital registry may represent patients who reach that hospital accurately while being a poor representation of every person in the community with the condition.

Bias is therefore tied to the estimand, target population, comparison and causal question.

4. Selection Bias Begins With Who Enters the Comparison

Selection bias occurs when the processes determining who or what enters, remains in or is analysed by the study create a distorted comparison for the question being asked.

It can begin before recruitment.

The sampling frame may omit a group.

It can begin during recruitment.

People with particular experiences may be more likely to volunteer.

It can begin after recruitment.

Participants with worse outcomes may be more likely to drop out.

Selection is a pipeline, not a single checkbox.

5. Sampling Bias Is One Form of Selection Problem

Suppose a school wants to know how students feel about homework and surveys only the students attending an optional after-school study session.

Those students may be unusually motivated, unusually supported or unusually concerned about achievement.

Even perfect measurement inside that group cannot make it automatically representative of the whole school.

How Sampling Works treats this doorway in full. The bias article owns the next question: how does that doorway systematically distort the estimate?

6. Volunteer Bias Appears When Participation Is Related to the Outcome

People who volunteer for a survey or study often differ from those who do not.

A health survey may attract more health-conscious respondents. A course evaluation may attract especially satisfied or dissatisfied students. An online political poll may attract people with strong views.

The danger is not simply that responders differ.

The danger is that the difference is connected to the quantity the researcher wants to estimate.

7. Attrition Bias Appears When Leaving Is Informative

A trial begins with two comparable groups.

Six months later, many participants have left.

If dropout is unrelated to treatment and outcome, the problem may mainly be lost precision.

If people leave because the treatment is ineffective, unpleasant or difficult, the observed completers can make the intervention look better than the experience of everyone originally assigned.

Attrition can therefore destroy the comparability created at the start.

8. Conditioning on a Collider Can Create Selection Bias

Some selection bias is more subtle than volunteer recruitment.

If two variables both affect whether a case enters the analysis, restricting to selected cases can make those variables appear associated even when they were not causally connected in the target population.

This is collider bias.

It is one reason variable roles must be mapped before adjustment or selection. The same variable that seems like a sensible “control” can open a non-causal path when it is a common effect of two other variables.

9. Information Bias Begins With How the World Is Recorded

Information bias occurs when exposure, outcome or other relevant variables are systematically measured or classified inaccurately.

The classic methodological families are often described as selection bias, information bias and confounding, with many named biases living inside those larger structures.

Information bias asks:

Did different parts of the world have different chances of being measured correctly?

10. Recall Bias Is Systematic Memory Error

Imagine asking people with and without a recent diagnosis to remember exposures from ten years ago.

Those with the diagnosis may search their memories more intensely, discuss possible causes with clinicians or reinterpret past events in light of current illness.

If exposure recall differs systematically by outcome status, the estimated exposure-outcome relationship can be distorted.

The lesson is not “memory is bad data”.

It is that the error process itself can depend on group membership.

11. Interviewer Bias Can Enter Through Human Measurement

An interviewer who knows a participant belongs to a particular group may unconsciously probe more deeply, paraphrase differently or interpret ambiguous answers differently.

Standardised scripts, training and blinding can reduce these pathways.

The key idea is that measurement instruments include people when people are part of the measurement process.

12. Observer and Detection Bias Can Favour the Expected Result

If outcome assessors know which treatment a participant received, ambiguous outcomes may be judged differently.

This is especially concerning for subjective outcomes such as ratings, imaging interpretation, behavioural coding or clinical judgement.

Blinding outcome assessors can close that information pathway where feasible.

But objective-looking measures can also contain judgement—algorithmic thresholds, coding decisions and device placement all embody choices.

13. Differential and Non-Differential Misclassification Behave Differently

Suppose a binary exposure is sometimes recorded incorrectly.

If the error pattern differs according to outcome or treatment group, the bias can move estimates in complicated directions.

If misclassification is similar across groups, it is often described as non-differential, but the familiar claim that such error “always biases toward the null” is too simple outside particular settings.

The direction of bias depends on the measurement and model structure.

14. Measurement Bias Can Come From the Instrument Itself

A sensor can drift. A scale can be calibrated incorrectly. A survey can use language understood differently across groups. A test can contain items that advantage students familiar with one cultural context.

Measurement bias therefore sits at the intersection of research bias and measurement validity.

A variable can be measured consistently and still systematically misrepresent the target construct.

15. Confounding Is a Causal Mixing Problem

Suppose people receiving Treatment A have worse baseline disease than people receiving Treatment B.

If the Treatment A group later has worse outcomes, part of the difference may reflect baseline severity rather than the treatment.

Treatment choice and outcome share a common cause.

That is confounding.

Cochrane describes confounding in intervention-effect estimates as arising when common causes influence both intervention assignment and outcome. Randomisation protects against this at assignment when implemented successfully.

16. Confounding by Indication Is Why Routine Treatment Comparisons Can Reverse the Story

In ordinary clinical care, sicker patients may be more likely to receive aggressive treatment.

If researchers simply compare treated and untreated patients, the treated group may have worse outcomes—not because treatment harms them, but because they started at higher risk.

The reason treatment was given becomes part of the causal structure.

The same pattern appears outside medicine. Students receiving intensive support may initially perform worse precisely because the weakest students were selected for support.

Observed outcome differences can therefore make helpful interventions look ineffective if baseline need is ignored.

17. Randomisation Targets Selection Bias at Assignment

Randomisation uses an unpredictable chance process to assign enrolled units to intervention groups.

When properly implemented, it prevents investigators from deliberately or accidentally directing particular prognostic types into one group and tends to balance both measured and unmeasured baseline factors in expectation.

But randomisation is not one word.

It requires sequence generation and allocation concealment.

If the next assignment can be predicted, recruiters may consciously or unconsciously influence who enters each group.

How Randomisation Works owns that mechanism in full.

18. Allocation Concealment and Blinding Attack Different Bias Pathways

These two are frequently confused.

Allocation concealment protects the assignment process before and until a participant is allocated. It prevents foreknowledge of the next assignment.

Blinding hides assigned intervention information after allocation from participants, clinicians, assessors or analysts where feasible.

CONSORT 2025 explicitly separates these jobs. Concealment targets selection bias at trial entry. Blinding can reduce performance or outcome-assessment bias after assignment.

19. Performance Bias Can Arise After Groups Are Assigned

Participants who know they received a novel intervention may change behaviour. Clinicians who know the assignment may provide different co-interventions. Teachers using a new curriculum may give extra enthusiasm or attention.

Now the difference between groups contains more than the nominal intervention.

Blinding can reduce some of these pathways in drug trials, but it is often impossible in education, surgery or public policy.

When blinding is impossible, designs need other protections: standardised co-interventions, objective outcomes, independent assessment, fidelity measurement and explicit interpretation of the remaining risk.

20. Detection Bias Concerns How Outcomes Are Determined

Suppose assessors know which patients received the experimental treatment.

If an outcome requires judgement, expectations may influence classification.

Blinded outcome assessment can reduce this risk.

But the importance depends on the outcome. All-cause mortality is less vulnerable to subjective assessment than a clinician-rated symptom scale, though even “objective” outcomes can be affected by differential surveillance or follow-up.

21. Surveillance Bias Appears When One Group Is Watched More Closely

A group receiving frequent medical follow-up may have more conditions detected than a group seen rarely.

A school programme that monitors participants weekly may detect more behavioural incidents simply because those students are observed more closely.

More detected events do not necessarily mean more underlying events.

The observation process becomes part of the outcome-generating system.

22. Lead-Time Bias Can Make Earlier Detection Look Like Longer Survival

Screening can detect a disease earlier without changing the time of death.

If survival is measured from diagnosis, the screened group appears to survive longer simply because the clock started earlier.

This is lead-time bias.

The repair is to use endpoints and designs that distinguish earlier detection from genuine improvement in outcomes.

23. Length Bias Can Make Screening Find Slower Disease Disproportionately

Diseases with longer detectable preclinical periods have more opportunities to be found by periodic screening.

Fast-progressing disease can appear between screening rounds and be underrepresented among screen-detected cases.

The detected group may therefore contain a larger proportion of naturally slower disease, making screening outcomes look better even before accounting for treatment benefit.

This is an example of how the sampling mechanism can select on disease dynamics.

24. Immortal-Time Bias Comes From Misclassified Time

In some observational studies, a participant must survive long enough to receive or qualify for an exposure.

If the period before exposure is incorrectly credited to the exposed group, that group contains time during which the outcome could not have occurred under the study’s classification.

This can create an artificial protective association.

The deeper lesson is that time itself is a research variable and can be misclassified.

25. Regression to the Mean Can Be Mistaken for Treatment Success

People are often selected for intervention because they have an extreme measurement: very high blood pressure, very low test score, unusually poor performance.

If part of that extreme value was temporary random fluctuation, the next measurement is likely to be less extreme even without an effective intervention.

A simple before-after comparison can therefore exaggerate improvement.

Appropriate comparison groups help reveal how much change would have occurred anyway.

26. History and Maturation Can Masquerade as Intervention Effects

A class improves after a new programme.

Perhaps the programme helped.

Students also grew older, practised other material, received ordinary teaching, sat mock examinations and experienced school-wide changes.

Without a suitable comparison, time-related changes can be attributed to the intervention.

This is why causal inference is a counterfactual problem rather than a before-after arithmetic problem.

27. Selective Outcome Reporting Changes Which Results Become Visible

A study can measure many outcomes but report only those that look favourable or interesting.

The published paper then presents a curated slice of the evidence-generating process.

Trial registration, protocols, preregistration and reporting standards create a record against which readers and reviewers can compare what was planned with what was reported.

The problem is not that exploratory outcomes exist.

The problem is losing the label that tells readers they were selected after looking.

28. Analysis Flexibility Can Produce Researcher Degrees of Freedom

Exclude one outlier or keep it?

Use one covariate set or another?

Transform the outcome?

Test one subgroup?

Stop data collection today or next week?

Many individually defensible choices can create a large analysis garden. If the final route is selected partly because it produces a preferred result and the search process is invisible, inferential error rates and effect estimates can become distorted.

Pre-specification, multiverse analysis, specification curves, held-out data and transparent reporting are different ways of making the analysis search more visible.

29. Optional Stopping Can Distort Statistical Evidence

Suppose researchers repeatedly test the data and stop collecting as soon as a conventional significance threshold is crossed.

If the statistical procedure does not account for that stopping rule, the chance of false-positive conclusions can rise.

Planned sequential designs can be valid because their decision rules are built into the analysis.

The issue is not looking at data per se. It is pretending the inferential procedure had one fixed sampling plan when the stopping rule actually depended on the evolving result.

30. Publication Bias Operates Above the Individual Study

Imagine ten equally careful studies.

Two find large positive effects.

Eight find small or uncertain effects.

If only the two striking studies are published, a literature search produces a misleading map even if those two papers are reported honestly.

Publication bias is therefore a selection problem at the level of the scientific record.

Systematic reviews must ask not only “What did the published studies find?” but “What studies and outcomes may be missing?”

31. Time-Lag Bias Can Make Exciting Results Arrive First

Studies with large or positive effects may sometimes be written, submitted and published more quickly than studies with null or less dramatic results.

During the early life of a research question, the visible literature can therefore be temporarily skewed toward excitement.

This is another reason early single-study headlines deserve restraint.

32. Citation Bias Can Distort Which Findings Become Influential

Positive, surprising or prestigious findings may be cited more often than less dramatic results.

Over time, a reader following citations can encounter a literature whose visibility is not proportional to evidential quality.

Citation preserves evidence lineage, but citation counts themselves are not neutral measures of truth.

33. Language Bias Can Hide Evidence Across Languages

A review limited to one language may miss relevant research published elsewhere.

If results differ systematically in which languages or journals they appear, the visible evidence base can be distorted.

Search strategy is therefore part of bias control in evidence synthesis.

34. Confirmation Bias Is Human, but Research Bias Needs Operational Specificity

Researchers, like everyone else, can prefer evidence that supports prior beliefs.

But saying “confirmation bias” is not enough to diagnose a study.

Where did the preference enter?

Hypothesis generation?
Selective recruitment?
Outcome choice?
Data exclusion?
Analysis selection?
Interpretation?
Publication?

Research methodology becomes useful when a broad psychological tendency is translated into a specific pathway that can be blocked or audited.

35. Conflict of Interest Is a Risk Factor, Not Proof of Bias

A financial or professional conflict can create incentives that raise concern about study design, interpretation or reporting.

Disclosure matters because readers need to know about those incentives.

But a conflict of interest does not logically prove that a result is wrong.

Likewise, absence of a declared financial conflict does not prove a study is unbiased.

The correct response is stronger scrutiny of the evidence pathway, not automatic acceptance or dismissal based on identity alone.

36. Risk of Bias Is Not the Same as a Quality Score

Modern evidence assessment increasingly avoids collapsing study quality into one crude numerical score.

Cochrane’s risk-of-bias approach evaluates domains tied to specific mechanisms, such as the randomisation process, deviations from intended interventions, missing outcome data, outcome measurement and selection of the reported result.

This is more informative than saying “Study A scored 7/10”.

Different biases can affect different outcomes in different directions.

Risk needs anatomy.

37. Bias Can Differ by Outcome Within the Same Study

An unblinded trial might have low risk of bias for mortality and higher risk for a subjective symptom score.

A missing-data pattern may affect a long-term outcome but not an immediate laboratory measurement.

This is why risk-of-bias judgements are often outcome-specific rather than one permanent label attached to the entire paper.

38. Prevention Is Usually Better Than Statistical Repair

Randomisation can prevent confounding at assignment more convincingly than modelling dozens of unmeasured baseline differences after the fact.

Blinded outcome assessment can prevent expectation from entering measurement more directly than trying to estimate the effect of assessor expectations later.

Prospective registration can preserve the original outcome plan more reliably than reconstructing intentions after publication.

Bias control is therefore an engineering problem:

Close dangerous pathways before they carry information.

39. Adjustment Can Reduce Confounding but Cannot Measure the Unmeasured Automatically

Observational studies often adjust statistically for measured confounders.

Regression, standardisation, matching, propensity scores and weighting are among the available approaches.

These can be powerful when the causal structure and measurements justify them.

But no method automatically adjusts for a confounder that was never measured or a variable whose measurement is badly wrong.

“Adjusted analysis” should never be translated mentally into “all bias removed”.

40. Sensitivity Analysis Asks How Strong Hidden Bias Would Need to Be

When bias cannot be ruled out, researchers can ask a more informative question than simply mentioning it in the limitations section.

How strong would unmeasured confounding need to be to erase the observed association?

How much outcome misclassification would change the conclusion?

How sensitive is the estimate to plausible missing-data mechanisms?

Quantitative bias analysis is an expanding methodological family built around such questions. A recent systematic review identified dozens of methods for using summary-level epidemiologic information to estimate or explain potential bias.

Sensitivity analysis does not make assumptions disappear.

It makes the consequence of assumptions more visible.

41. Negative Controls Can Reveal Hidden Structure

A negative-control exposure or outcome is chosen because it should not be causally affected in the hypothesised way but may share some bias pathways.

If an association appears where no causal effect is expected, hidden confounding or measurement structure may be operating.

Negative controls do not solve every causal problem, but they give the world another opportunity to expose a false pathway.

42. Triangulation Works When Methods Fail Differently

If several studies use the same biased measurement, agreement among them may simply reproduce the same distortion.

Triangulation becomes stronger when different methods have different likely biases.

A randomised trial, natural experiment, longitudinal cohort and mechanistic laboratory study may each see a different part of the problem.

Convergence is most informative when the bias structures are not perfectly shared.

43. Replication Can Reveal Bias Boundaries

A finding that appears only in one laboratory, one dataset, one measurement or one population may depend on a hidden local pathway.

Independent replication changes people, samples and sometimes methods.

If the effect survives, some local-bias explanations become less plausible.

If it fails, investigators should not immediately declare the first study fraudulent or the second incompetent. The disagreement becomes a map of where hidden bias, measurement differences or true context dependence may live.

44. Peer Review Can Detect Bias but Cannot Guarantee Its Absence

Reviewers can identify obvious selection problems, weak blinding, inappropriate adjustment, missing outcomes and overclaiming.

But reviewers work from what the manuscript and available materials make visible.

If an unreported outcome was never mentioned, a reviewer may not know it existed. If data fabrication is convincing, ordinary review may not detect it.

Peer review is one bias-control checkpoint, not the owner of all error detection.

45. Systematic Reviews Can Inherit Bias From Every Included Study

Meta-analysis does not wash bad studies clean by averaging them.

If constituent studies share unmeasured confounding, selective reporting or measurement problems, a pooled estimate can be precise and misleading.

Methodological guidance therefore recommends assessing study-level risk of bias and using sensitivity analyses where relevant rather than treating sample size as the sole evidence-strength metric.

46. AI and Big Data Create New Scales for Old Bias Problems

An AI model can train on millions or billions of records.

If those records systematically overrepresent some populations, under-record certain outcomes or inherit past institutional decisions, the model can reproduce those structures.

A benchmark can also be biased toward tasks that are easy to score rather than tasks that matter in deployment.

Scale multiplies exposure to the data-generating process.

It does not make that process neutral.

47. Algorithmic Bias Is Not One Mechanism

An algorithm can produce unequal or inaccurate outcomes for many reasons:

biased sampling;
historical labels;
measurement proxies;
objective-function choices;
threshold selection;
distribution shift;
different error costs across groups;
or feedback loops after deployment.

Calling all of this simply “algorithmic bias” can hide the repair.

Research-bias thinking asks for the specific distortion pathway.

48. Education Has a Selection-Bias Trap Called “The Students Who Need Help”

Suppose a school gives extra tutoring to the weakest students.

At the end of term, tutored students still score lower than untutored students.

It would be absurd to conclude automatically that tutoring caused lower performance.

The programme was allocated because students were already weaker.

This is confounding by need.

Education is full of such targeted interventions. The receiver-selection rule must be reconstructed before outcome differences can be interpreted.

49. Education Also Has Survivorship Bias

A school celebrates study habits used by top-performing students.

Perhaps those habits helped.

Perhaps students with strong prior knowledge, stable routines and high motivation found those habits easier to sustain.

Looking only at successful survivors hides students who used the same strategy and struggled or students who succeeded by a different route.

Anecdotes from winners are a selected sample of pathways.

50. Journalism Can Amplify Research Bias Through Headline Selection

Newsrooms understandably prefer surprising findings.

A nuanced replication, careful null result or methodological correction may receive less attention than a dramatic first study.

The public evidence environment can therefore become more biased than the scientific literature it draws from.

Research literacy requires reading past the novelty filter and asking how the result sits inside the cumulative evidence base.

51. The Hostile Test: One Million App Users

A company analyses one million users and finds that people using Feature X achieve better outcomes.

The confidence interval is tiny.

But Feature X was optional.

Users who choose it are more experienced, more engaged and more likely to complete difficult tasks.

The million observations give a very precise association.

They do not by themselves tell us what would happen if otherwise comparable users were assigned the feature.

Precision cannot manufacture a counterfactual.

52. The Second Hostile Test: Perfect Randomisation, Biased Outcome Measurement

A trial randomises participants flawlessly.

Then unblinded assessors score the treatment group more generously on a subjective outcome.

Randomisation solved baseline confounding.

It did not solve biased outcome ascertainment.

This is why “RCT” is a design family, not a magic quality badge.

53. The Third Hostile Test: The Meta-Analysis of Only Visible Studies

Twenty small experiments are conducted.

Five striking positive studies are published quickly.

Several null studies remain in drawers or conference archives.

A meta-analysis of the published papers finds a convincing pooled effect.

The arithmetic is correct for the visible studies.

The evidence sample is biased.

Bias can therefore occur at the level of selecting studies just as it occurs when selecting people.

54. The Fourth Hostile Test: Bias Correction That Creates New Bias

An analyst worries about confounding and adjusts for every available variable.

One variable is a collider.

Another is a mediator.

Now the “adjusted” estimate is farther from the desired causal effect.

Bias control is not synonymous with maximum adjustment.

It requires the right causal map.

55. Primary School: Bias Begins as “Did We Give Everything a Fair Chance?”

Young learners can understand the foundation without methodological vocabulary.

If you want to know which seed grows fastest, do not give one seed sunlight and the other darkness.

If you want to know what the class thinks, do not ask only your three best friends.

If you know which cup is supposed to contain the “better” drink, do not score taste differently because you expect it to win.

The child learns the core idea:

A fair test must stop my method from pushing the answer toward the result I expect.

56. Secondary School: Separate Random Error From Systematic Error

Secondary students can learn that repeating measurements can reduce the influence of random fluctuation but cannot automatically repair a systematically miscalibrated instrument or unfair comparison.

They can identify simple selection, observer and measurement biases and begin asking how controls, randomisation and blinding change the evidence pathway.

57. JC and University: Bias Becomes a Causal Architecture

At higher levels, “this study may be biased” is too vague.

Students should identify:

the target estimand;
the mechanism creating distortion;
the likely direction where knowable;
the design stage where it enters;
the prevention or adjustment available;
the assumptions required for repair;
and the residual uncertainty after repair.

Bias stops being a criticism word and becomes an engineering diagram of failure.

58. Where Research Bias Fits in the eduKateSG “How Works” Landscape

How Scientific Research Works — the whole evidence-producing route.
How Research Variables Work — the causal roles through which confounding and collider bias are understood.
How Sampling Works — selection into the study.
How Measurement Works — information quality and systematic measurement error.
How Peer Review Works — one external scrutiny layer.
How Uncertainty Works — what remains unresolved after bias controls.
How Verification Works — checking claims against specified evidence.

Bias is the distortion map that asks where those otherwise useful mechanisms can systematically bend the result.

59. What This Article Does Not Claim

Every imperfection is not automatically bias.
Large samples reduce some random error but do not automatically remove systematic bias.
A conflict of interest is not proof that a result is false.
Randomisation controls important bias at assignment but not every later bias.
Blinding and allocation concealment solve different problems.
Statistical adjustment cannot guarantee removal of unmeasured confounding.
Adding every available covariate can create new bias.
Risk-of-bias assessment is not a universal numerical quality score.
A single bias label rarely tells you the direction or magnitude of distortion.
Sensitivity analysis explores robustness under assumptions; it does not create information the study never collected.

60. A Compact Bias Audit

What is the exact target quantity or causal effect?
Who could enter the study?
Who could not?
Who volunteered?
Who dropped out?
Could selection depend on exposure and outcome?
How were exposure and outcome measured?
Could measurement differ between groups?
Were assessors blinded where feasible?
What common causes could confound the comparison?
Was treatment assignment randomised?
Was the allocation sequence concealed?
Were post-treatment variables adjusted for?
Could any adjusted variable be a collider?
Were outcomes and analyses pre-specified?
How many analytic routes were searched?
Were null or unfavourable outcomes reported?
Could unpublished studies be missing?
What sensitivity analyses test plausible remaining bias?
Has independent replication challenged the same result through a different bias structure?

61. Frequently Asked Questions

What is research bias?

Research bias is systematic distortion in study design, conduct, measurement, analysis, reporting or publication that causes an estimate or evidence base to differ from the target it is intended to represent.

What is the difference between bias and random error?

Random error causes estimates to fluctuate because of finite or noisy observations. Bias pushes results systematically because of the way units are selected, measured, compared, analysed or reported.

Does a larger sample remove bias?

Not automatically. A larger sample can make a biased estimate more precise if the same systematic distortion remains.

What are the main families of research bias?

Common high-level families include selection bias, information or measurement bias, confounding and reporting or publication biases, with many more specific named mechanisms inside them.

Can bias be corrected statistically?

Some bias can be reduced or explored using appropriate adjustment, weighting, missing-data methods or quantitative sensitivity analysis, but these methods depend on assumptions and cannot reliably reconstruct every unmeasured or unrecorded process.

62. Authoritative Research Corridor

Final Thought: Bias Is a Bent Route, Not a Bad Person

It is tempting to make bias a moral accusation.

Sometimes misconduct is real and must be treated seriously.

But much research bias can occur without anyone intending to deceive.

A recruiter knows the next allocation.

A survey misses people who are hardest to reach.

An assessor knows who received treatment.

An outcome is easier to publish when it is exciting.

A confounder was never measured because nobody realised it mattered.

The scientific response is therefore not merely to demand better intentions.

Build better routes.

Hide future allocations.

Blind measurements where possible.

Preserve missing cases.

Pre-specify important decisions.

Publish inconvenient results.

Test sensitivity.

Invite replication.

A trustworthy research system does not assume humans are unbiased. It designs so that bias has fewer places to hide.

How Research Bias Works | From Selection, Measurement and Confounding to Reporting Distortion, Risk-of-Bias Control and Better Inference