Skip to content
Scintria
ToolsGrant writerFunding
HomeLiteratureClinical trialsEvidence LabProtocol packData analysisLearnProjectsLibraryReferencesApplications
Start free
ScintriaStart free

Learn / Epidemiology / Bias, confounding and causation

Selection bias and information bias

14 minute read · 6-question quiz

Download this lesson as a colour Word or PDF study guide with the figures and answers. Included in paid plans; reading and quizzes stay free.Sign in

What you will learn

  • Define systematic error and distinguish bias from random error and confounding

  • Recognise common forms of selection bias and information bias

  • Show how non-differential and differential misclassification change an odds ratio

Every study estimate has two kinds of error. Random error is chance and shrinks with sample size. Systematic error (bias) is a flaw in the design, conduct or analysis that pushes the estimate in a consistent direction. Bias cannot be fixed with more participants and cannot be measured by a p-value.

Three broad families:

  • Selection bias: who gets into the study, or the analysis, distorts the comparison.
  • Information (measurement) bias: exposure, outcome or covariates are measured wrongly.
  • Confounding: a third factor mixes up the effect (next lesson).

Selection bias

Volunteer biasBerkson (hospital) biasLoss to follow-upPeople who volunteer arehealthier or moremotivatedExample: people joining ascreening trialHospital patients differfrom the communityTwo diseases can looklinked because each raisesthe chance of admissionThose who drop out differfrom those who stayCohorts and trials
Selection bias (1 of 2): who gets into the study.
Healthy-worker effectNon-responseSurvivor biasWorkers are healthier thanthe general populationOccupational studiesNon-responders differ fromrespondersSurveysOnly survivors are studiedPrevalent cases incase-control studies
Selection bias arises when the people analysed differ systematically from the target population, or when the groups compared differ in ways other than exposure.

Information bias

Information bias is misclassification: people placed in the wrong exposure or outcome category.

  • Non-differential misclassification: error is the same in all comparison groups (for example both cases and controls have a laboratory test with 90% specificity). For a binary exposure it usually pushes the estimate toward the null: a true association looks weaker.
  • Differential misclassification: error differs by group (cases remember exposures more completely than controls). It can bias in either direction.

Common sources:

  • Recall bias: mothers of children with a birth defect search their memory for exposures; mothers of healthy children do not.
  • Interviewer (observer) bias: the interviewer knows who is a case and probes more.
  • Social desirability bias: under-reporting of stigmatised behaviour (sex work, drug use, non-adherence).
  • Detection (surveillance) bias: an exposed group is examined more carefully, so more disease is found.
  • Instrument error: miscalibrated scales, poor-quality tests, inconsistent case definitions.

Worked example: misclassification and the odds ratio

A case-control study with 100 cases and 200 controls, true exposure:

CasesControls
Truly exposed10050
Truly unexposed100150

True OR = (100 × 150)/(100 × 50) = 3.0.

Non-differential error. Exposure is measured by a questionnaire with sensitivity 80% and specificity 90% in both groups.

  • Cases: apparent exposed = 100 × 0.8 + 100 × 0.1 = 90; unexposed 110.
  • Controls: apparent exposed = 50 × 0.8 + 150 × 0.1 = 55; unexposed 145.
  • Observed OR = (90 × 145)/(110 × 55) = 2.16. The effect is diluted from 3.0 toward 1.

Differential error (recall bias). Cases report with sensitivity 95% and specificity 90%; controls with sensitivity 70% and specificity 95%.

  • Cases: apparent exposed = 105; unexposed 95. Controls: apparent exposed = 42.5; unexposed 157.5.
  • Observed OR = (105 × 157.5)/(95 × 42.5) = 4.1. The effect is exaggerated.
01234Odds ratio3True OR2.16Non-differential error4.1Differential (recall) error
The same data seen through different measurement errors.

Prevention is better than correction

Bias is best handled at the design stage: precise case definitions, objective measures (laboratory tests, records, biomarkers), standardised and blinded data collection, high response rates, careful choice of controls, and complete follow-up. Analytical fixes (sensitivity analysis, bias adjustment, inverse-probability weights) help but rest on assumptions.

  1. 1

    Ask: who is in the study?

    Define the target population and compare it with those eligible, invited, agreed, completed and analysed. Each gap is a place for selection bias.

  2. 2

    Ask: how was it measured?

    Would the measurement be equally accurate in all groups? Did anyone know the hypothesis or the group?

  3. 3

    Ask: which direction?

    Predict whether the bias would push the estimate toward or away from the null, and how much.

  4. 4

    Ask: could it explain the whole finding?

    A bias that would need to be implausibly large to explain the result is a minor worry.

Bias is a property of the study, not the result

A statistically significant, narrow-CI result can still be wrong. Report limitations honestly and discuss the likely direction of each bias, as STROBE and CONSORT require.

Summary

  • Bias is systematic error; random error shrinks with sample size but bias does not.

  • Selection bias concerns who is studied; information bias concerns how they are measured.

  • Non-differential misclassification of a binary exposure usually biases toward the null; differential error can go either way.

  • Recall, interviewer, detection and social desirability biases are common information biases.

  • Prevent bias in design; sensitivity analyses can explore the rest.

Check your understanding

Question 1
1

A cohort in which sicker participants drop out has:

Question 2
2

Non-differential misclassification of exposure usually:

Question 3
3

Mothers of children with leukaemia recall chemical exposure better than control mothers. This is:

Question 4
4

Women using contraceptives are screened more often and so more breast lesions are found in them. This is:

Question 5
5

A bigger sample size will reduce:

Question 6
6

Objective measurements such as laboratory tests mainly help reduce:

Draft study guide. Spotted an error or something unclear? Tell us at hello@scintria.com. This is educational material, not clinical or regulatory advice.

Mortality, life expectancy and the burden of disease (DALYs)Confounding, mediators, colliders and causal diagrams