Information Bias

How information bias distorts an estimated relationship

Information bias occurs when exposure, outcome, covariate, or eligibility information is measured, recorded, classified, or retrieved inaccurately in a way that distorts an association. It can arise in primary data collection, routine records, linkage, self-report, clinical assessment, coding, or algorithms. This page explains major forms of information bias, their likely direction, and the design and analysis methods used to prevent or assess them.

Information bias is systematic measurement error

Random measurement error adds variability, while information bias systematically changes group comparisons or effect estimates. The same measurement problem can be random in one setting and differential in another. Direction cannot be assumed without understanding how error relates to exposure, outcome, and other variables.

Error type	Definition	Typical consequence
Non-differential misclassification	Error does not differ by the variable being compared under the stated conditions	Often attenuates simple binary associations, but not always
Differential misclassification	Error differs by exposure, outcome, group, or another relevant factor	Can bias in any direction
Random continuous error	Observed values vary around the true value	Often reduces precision and can attenuate regression slopes
Systematic continuous error	Values are consistently shifted or scaled	Changes means, thresholds, and effect estimates
Missing information	Measurement is absent in a non-random pattern	Can create selection and information problems

Misclassification changes categorical variables

Misclassification occurs when a person is placed in the wrong exposure, outcome, or covariate category. Its effect depends on sensitivity, specificity, prevalence, number of categories, and whether error differs across comparison groups. The statement that non-differential error always biases toward the null is not generally true.

For a binary measure:

$$ Sensitivity=\frac{True\ positives}{True\ positives+False\ negatives} $$

$$ Specificity=\frac{True\ negatives}{True\ negatives+False\ positives} $$

These parameters should refer to a credible reference standard and the population in which the measure is used.

Recall bias depends on memory and group status

Recall bias occurs when past exposure or events are remembered or reported differently according to outcome status or another study characteristic. People with a diagnosis may search their memory more intensely than controls, while long recall periods can reduce accuracy for everyone. Emotion, salience, stigma, and interviewer prompts can change reporting.

Prevention can include prospective measurement, records collected before outcome occurrence, bounded recall periods, memory aids, neutral questions, and blinding interviewers to study hypotheses. Validation subsamples can estimate the magnitude and direction of error.

Interviewer and observer bias affect recorded information

An interviewer or assessor who knows exposure, treatment, or outcome can probe, interpret, or record information differently. The problem can affect clinical ratings, diagnostic assessment, chart review, and qualitative coding. Standardised instruments do not eliminate bias when their use still requires judgment.

Training, manuals, blinded assessment, inter-rater reliability, central adjudication, audio review, and automated prompts can reduce variation. Blinding feasibility and success should be reported rather than assumed.

Detection and surveillance bias change outcome ascertainment

People receiving more monitoring have more opportunities for an outcome to be detected. A treatment group with frequent laboratory testing may appear to have more abnormalities even when underlying incidence is similar. Diagnostic suspicion can also lead to differential testing.

Analyses should examine testing frequency, follow-up intensity, healthcare contact, and outcome definitions. Restricting to objectively severe outcomes, standardising surveillance, or adjusting for observation processes may help, but adjustment can introduce bias if healthcare contact lies on the causal pathway.

Diagnostic suspicion can create circular evidence

Knowledge of exposure can influence the likelihood of assigning a diagnosis, and knowledge of diagnosis can influence how past exposure is classified. This is especially problematic when diagnostic criteria include the exposure or when clinical notes repeat unverified assumptions. Independent and blinded adjudication can break the circular pathway.

The reference standard should not incorporate the index measure being evaluated unless the implications are explicit. Incorporation bias can exaggerate apparent accuracy.

Self-report can be affected by social desirability

Respondents may underreport stigmatised behaviour and overreport behaviour perceived as desirable. Privacy, question wording, interviewer presence, legal risk, and cultural context affect the direction and magnitude. The problem is not evidence that respondents are intentionally deceptive; the measurement context shapes disclosure.

Confidential self-administration, indirect questioning, validated instruments, biomarkers, and record linkage can improve measurement. Alternative sources can have their own error and should not be treated as perfect automatically.

Routine health records contain measurement processes

Electronic health records and claims are generated for care and payment rather than research. Diagnosis codes can represent confirmed disease, suspected disease, rule-out testing, history, or billing. Absence of a code can mean absence of disease, no contact, incomplete recording, or care outside the system.

Phenotypes should define code sets, time windows, care settings, repeat requirements, laboratory thresholds, medication evidence, and exclusions. Validation should report positive and negative predictive performance when possible.

Exposure measurement can vary over time

Medication, behaviour, occupation, and environmental exposures can change. A baseline measure may not represent follow-up, while future information must not be used to classify earlier exposure. Time-varying measurement and appropriate risk windows are needed when current exposure determines risk.

Prescription, dispensing, possession, administration, and ingestion are different exposure concepts. Choosing one as a proxy for another introduces error whose direction depends on adherence and data capture.

Outcome measurement can be subjective or objective

Objective instruments can reduce some observer effects but remain vulnerable to calibration, threshold, timing, device, specimen, and processing error. Patient-reported outcomes directly measure symptoms but can be influenced by recall, framing, expectations, and missingness. The appropriate measure depends on the concept, not on a simple subjective-objective hierarchy.

Outcome definitions should specify who assessed, which instrument, timing, threshold, adjudication, and competing events. Differential follow-up or treatment awareness should be considered explicitly.

Covariate error can leave residual confounding

An analysis may adjust for a confounder measured with error and still retain confounding. Crude categories can also fail to capture nonlinear or time-varying relationships. A long list of adjusted variables does not guarantee adequate control.

Validation data, repeated measurements, calibration models, regression calibration, simulation-extrapolation, or probabilistic bias analysis can assess the effect. Correction requires assumptions about the measurement-error process.

Continuous measurement error can attenuate associations

In a simple linear regression with classical random error in exposure, the observed slope is often attenuated toward zero. If (X^*=X+u), with error (u) independent of the true exposure and outcome model, the attenuation factor is:

$$ \lambda=\frac{Var(X)}{Var(X)+Var(u)} $$

The observed coefficient is approximately:

$$ \beta_{observed}=\lambda\beta_{true} $$

These results depend on restrictive assumptions. Differential, systematic, Berkson, correlated, or outcome measurement error can behave differently.

Differential error can bias in any direction

When measurement accuracy differs by study group, the effect estimate can move toward or away from the null or even reverse. For example, an unblinded assessor may rate treatment-group symptoms more favourably, while cases may report prior exposures more completely than controls. Direction should be reasoned from the measurement pathway rather than asserted from a rule of thumb.

Bias diagrams, sensitivity parameters, and corrected two-by-two tables can make assumptions explicit. Several plausible error scenarios should be tested when validation data are unavailable.

Linkage error creates false and missed matches

Record linkage can assign records to the wrong person or fail to link records belonging to the same person. Error may differ by name changes, migration, housing instability, ethnicity, age, or data quality, creating both bias and inequity. Deterministic and probabilistic linkage require documented thresholds and evaluation.

Analysts should report match rates, false-match and missed-match estimates, clerical review, linkage variables, blocking, and subgroup performance. Treating linked data as error-free can make large datasets falsely reassuring.

Algorithmic classification is still measurement

Natural-language processing, machine learning, and rule-based phenotypes convert raw records into analytical variables. Their sensitivity, specificity, calibration, and transportability determine information quality. Performance can deteriorate when documentation practice, population, software, or disease prevalence changes.

Validation should use a representative reference sample and report subgroup performance. A high overall area under the curve does not guarantee accurate classification at the chosen threshold or equal error across groups.

Missing data and information bias can interact

Missingness is not identical to measurement error, but both can arise from the same process. Patients with severe illness may have more complete clinical data, while people with access barriers may be absent from records. Imputation fills missing values under assumptions; it does not correct systematically wrong observed values.

The analysis should separate unavailable, unmeasured, not applicable, below detection, and structurally missing data. Coding all as zero creates misclassification.

Validation studies estimate measurement performance

A validation study compares the operational measure with a stronger reference method in a relevant sample. Verification should not be restricted only to positive cases because sensitivity and specificity then cannot both be estimated. The reference standard can also be imperfect and may require latent-class or composite approaches.

The validation sample should represent the range of values, settings, and groups in the main study. Transporting measurement parameters from another context requires justification.

Quantitative bias analysis makes assumptions visible

Probabilistic bias analysis specifies distributions for sensitivity, specificity, or other error parameters and propagates them to corrected effect estimates. It does not remove bias automatically; it shows how conclusions change under stated assumptions. Correlation and differential error should be represented where relevant.

For observed prevalence (p_{obs}) under non-differential binary misclassification, one correction is:

$$ p_{true}=\frac{p_{obs}+Sp-1}{Se+Sp-1} $$

The denominator must be positive and the assumptions must fit the setting. Sampling uncertainty and uncertainty in (Se) and (Sp) should be propagated.

Multiple sources can improve or complicate measurement

Combining self-report, clinical records, registries, laboratory data, and devices can improve ascertainment when sources capture complementary information. It can also create conflicting definitions and differential availability. A hierarchy, adjudication rule, or latent measurement model should be specified before results are known.

Agreement statistics alone do not identify which source is correct. Disagreement should be examined for systematic patterns and subgroup differences.

Information bias affects evidence synthesis

Studies can use different exposure and outcome definitions, measurement instruments, thresholds, and ascertainment intensity. Pooling them as though they measure the same construct can create heterogeneity or a misleading average. Reviewers should extract measurement methods and assess risk of bias at the outcome level.

Sensitivity analyses can restrict to validated measures, blinded outcomes, common definitions, or low-risk studies. Meta-regression may explore measurement differences but is limited when few studies are available.

Information bias affects economic models

Misclassified disease, treatment, response, adverse events, or resource use can distort model inputs and state membership. Measurement error can influence both mean parameters and correlations. Model uncertainty should include plausible corrected values or scenarios when the input evidence is vulnerable.

Examples include:

• Claims codes that miss disease managed outside hospital.
• Self-reported adherence that overstates doses taken.
• Unblinded response measures that exaggerate benefit.
• Incomplete adverse-event reporting that understates costs and disutility.
• Resource-use records that omit informal or out-of-system care.

Worked misclassification example

Suppose a screening measure has sensitivity 80% and specificity 90% in a population where true prevalence is 10%. Among 1,000 people, it identifies 80 true positives and 90 false positives. Its positive predictive value is therefore:

$$ PPV=\frac{80}{80+90}\approx47.1% $$

More than half of positive classifications are false despite apparently good sensitivity and specificity. The result depends on prevalence and shows why classification performance must be evaluated in the intended population.

Prevention begins in study design

Analysis cannot fully repair information that was never measured well. Prevention includes clear operational definitions, validated instruments, training, calibration, blinding, standard timing, neutral questioning, repeated measures, data-quality monitoring, and accessible participation. Pilot work should test both accuracy and feasibility.

The measurement protocol should preserve raw values, units, dates, provenance, changes, and reasons for correction. Derived classifications can then be reproduced when definitions change.

Common mistakes

Information bias is often invoked without specifying which variable is wrong, how error differs, or which direction is plausible. The following errors weaken both diagnosis and correction. Each should be checked before attributing a result to measurement bias.

• Assuming all non-differential misclassification biases toward the null.
• Treating routine records or biomarkers as error-free because they are objective.
• Using a reference standard that incorporates the index measure.
• Estimating PPV without considering prevalence.
• Adjusting for a mismeasured confounder and assuming confounding is removed.
• Treating imputation as correction of inaccurate observed data.
• Applying validation parameters from a different population without justification.
• Reporting algorithm discrimination without threshold-specific errors.
• Ignoring subgroup differences in linkage or classification performance.
• Naming recall bias simply because exposure was self-reported.

Reporting information bias

Transparent reporting should describe the complete measurement pathway and identify where error could enter. It should report validation evidence, likely direction, and sensitivity of conclusions rather than asserting that bias is possible in general. Corrected estimates should remain linked to their assumptions.

• Define each exposure, outcome, covariate, and classification threshold.
• State the source, respondent, assessor, instrument, timing, and blinding.
• Report missingness, repeat measurements, calibration, and data corrections.
• Report sensitivity, specificity, predictive values, reliability, and reference standard where relevant.
• Explain whether error is expected to be differential and why.
• Report validation-sample representativeness and subgroup performance.
• Present quantitative bias or sensitivity analyses with parameter sources and uncertainty.
• State how measurement limitations affect the conclusion and economic model inputs.

The decision standard

Information bias is decision-relevant when measurement or classification error systematically changes the estimated relationship or the population to which it applies. A credible assessment names the affected variable, reconstructs the measurement process, and evaluates direction and magnitude rather than relying on a generic label. The strongest protection comes from valid measurement design, representative validation, and analyses that propagate remaining uncertainty transparently.