Written and maintained by CASRAI Editorial Board
Last updated
What is information bias?
Information bias is systematic error that arises when the information used in a study is measured, recorded or collected inaccurately. The Catalogue of Bias defines it as a systematic difference from the truth that arises in the collection, recall, recording and handling of information in a study. Kesmodel (2018) puts it as occurring when any information used in a study is measured or recorded inaccurately. It can affect the exposure, the outcome or a confounder, and it is a property of how data were gathered, not of who was enrolled. That separates it from selection bias, which concerns how participants got into the study.
The Catalogue of Bias names four major types: misclassification bias, observer bias, recall bias and reporting bias. This page defines the concept, shows with arithmetic how measurement error changes an odds ratio, and explains what to do about it at the design stage.
Information bias examples
A published example is a study of young adults’ computer use (Chang et al., 2010, as summarised by the Catalogue of Bias). Participants logged their daily computer time for three weeks while software on their computers measured actual use. For reference-estimated use below 3.6 hours a day, self-reports tended to overestimate the duration of use; above 3.6 hours a day, they tended to underestimate it. The measurement error came from the self-report instrument, so the exposure was misclassified.
A second example is recall. The Catalogue of Bias notes that parents of children diagnosed with cancer may be more likely to recall infections earlier in the child’s life, which can create a false or partly false link between infection and cancer. Our guide to recall bias covers it in depth.
Types of information bias and how each is reduced
| Type | What it is | Reducing it (as stated by the source) |
|---|---|---|
| Misclassification bias | Individuals are assigned to a different category than the one they should be in | Use the most accurate measurements available and think carefully about how individuals or data points are categorised |
| Recall bias | Systematic error due to differences in accuracy or completeness of recall of past events or experiences | A prospective design; neutral questioning in case-control studies so that wording does not shape the answer |
| Observer bias | Systematic discrepancies from the truth in the process of observing and recording information; see observer bias | Blind outcome assessors to exposure status; train observers in how to record findings |
| Reporting bias | Systematic distortion from the selective disclosure or withholding of information by parties involved in a study | Transparency: prospective trial registration, sharing data and code, and reporting guidelines such as CONSORT |
Differential and non-differential misclassification
The Catalogue of Bias distinguishes two forms. In non-differential misclassification, “the probability of individuals being misclassified is equal across all groups in the study”. In differential misclassification, “the probability of being misclassified differs between groups”. Kesmodel (2018) reports that non-differential exposure misclassification in cohort studies usually biases results toward no association, whereas differential misclassification can bias in either direction, especially when exposure is recorded after the outcome. Our comparison of differential vs non-differential misclassification sets the two side by side.
Formula: how sensitivity and specificity produce the observed counts
Information bias has no single formula. What can be computed is how a known error rate changes a result. For a binary exposure, the expected observed number of exposed people in a group is:
Observed exposed = (true exposed × Se) + (true unexposed × (1 − Sp))
- Se (sensitivity): the proportion of truly exposed people classified as exposed.
- Sp (specificity): the proportion of truly unexposed people classified as unexposed, so 1 − Sp is the share of unexposed people wrongly counted as exposed.
- Observed unexposed = group total − observed exposed.
- OR = (a × d) / (b × c), where a and b are the exposed and unexposed cases, and c and d are the exposed and unexposed controls.
Worked example with real numbers
This is an illustrative hypothetical with assumed numbers, not data from a real study. A case-control study has 1,000 cases and 1,000 controls. True exposure: 400 exposed and 600 unexposed cases; 200 exposed and 800 unexposed controls.
True odds ratio = (400 × 800) / (600 × 200) = 320,000 / 120,000 = 2.67.
Scenario A, non-differential error (Se = 0.80 and Sp = 0.90 in both groups).
- Cases observed exposed = 400 × 0.80 + 600 × 0.10 = 320 + 60 = 380; unexposed = 620.
- Controls observed exposed = 200 × 0.80 + 800 × 0.10 = 160 + 80 = 240; unexposed = 760.
- Observed OR = (380 × 760) / (620 × 240) = 288,800 / 148,800 = 1.94.
Equal error in both groups moved the odds ratio from 2.67 toward 1 (the null).
Scenario B, differential error (cases Se = 0.95, Sp = 0.85; controls Se = 0.80, Sp = 0.95).
- Cases observed exposed = 400 × 0.95 + 600 × 0.15 = 380 + 90 = 470; unexposed = 530.
- Controls observed exposed = 200 × 0.80 + 800 × 0.05 = 160 + 40 = 200; unexposed = 800.
- Observed OR = (470 × 800) / (530 × 200) = 376,000 / 106,000 = 3.55.
Here the error differs by group and the odds ratio moved away from the null, to 3.55. With other error rates it could move the other way, which is why differential error is harder to anticipate.
Does information bias push results toward the null?
Not always. The table summarises what the example and Kesmodel (2018) show.
| Pattern | Typical effect on the estimate |
|---|---|
| Non-differential exposure misclassification, cohort study | Usually toward no association (Kesmodel, 2018); in Scenario A, OR 2.67 became 1.94 |
| Differential misclassification | Either direction (Kesmodel, 2018); in Scenario B, OR 2.67 became 3.55 |
| Misclassified confounders or disease status | Can also bias results (Kesmodel, 2018) |
Assumptions and common mistakes
- The arithmetic above assumes a binary exposure and known, fixed sensitivity and specificity. Real error rates are estimated, and they vary.
- Treating “non-differential” as harmless. It describes equal error across groups; it still changes the estimate.
- Assuming the bias always shrinks the effect. Differential error can inflate or reverse it.
- Relying on analysis alone. Kesmodel (2018) states that information bias is probably best prevented during planning of data collection, as few and insufficient methods exist for correcting inaccurate information. For sensitivity analyses see quantitative bias analysis.
How to report information bias
Describe how each key variable was measured, who recorded it and when, and which direction the likely error would push the estimate. An example sentence: “Exposure was self-reported; if misclassification was equal across groups, the observed odds ratio would be expected to lie closer to 1 than the true value.” In APA style, give the estimate as OR = 1.94 with its 95% confidence interval, and name the source of any sensitivity and specificity values used. State any validation data you used. Related guides: response bias and social desirability bias.
How to calculate it in Excel, R or Python
These reproduce Scenario A. Replace the numbers with your own assumed error rates.
Excel (A1 = true exposed cases 400, B1 = true unexposed cases 600, C1 = true exposed controls 200, D1 = true unexposed controls 800):
E1: =A1*0.8+B1*(1-0.9)
F1: =(A1+B1)-E1
G1: =C1*0.8+D1*(1-0.9)
H1: =(C1+D1)-G1
I1: =(E1*H1)/(F1*G1) R:
obs_exposed <- function(exp, unexp, se, sp) exp*se + unexp*(1-sp)
ae <- obs_exposed(400, 600, 0.8, 0.9); ce <- obs_exposed(200, 800, 0.8, 0.9)
or_obs <- (ae * (1000 - ce)) / ((1000 - ae) * ce)
round(or_obs, 2) # 1.94Python:
def obs_exposed(exp, unexp, se, sp):
return exp * se + unexp * (1 - sp)
ae = obs_exposed(400, 600, 0.8, 0.9) # 380
ce = obs_exposed(200, 800, 0.8, 0.9) # 240
print(round((ae * (1000 - ce)) / ((1000 - ae) * ce), 2)) # 1.94How to prevent information bias
Use the most accurate measurements available, categorise carefully, and plan data collection before the study starts. For recall bias, the Catalogue of Bias calls a prospective design the most appropriate way to avoid it and advises neutral questioning in case-control studies. For related design questions see research methods.
FAQ
What is information bias?
It is systematic error from inaccurate measurement, recording or collection of study information. It includes misclassification, recall, observer and reporting bias.
What are examples of information bias?
Self-reported computer use that differed from software-monitored use (Chang et al., 2010), and parents of children with cancer recalling earlier infections more readily (recall bias).
Is information bias the same as selection bias?
No. Information bias concerns errors in collecting or recording data; selection bias concerns how participants entered the study.
Can information bias be corrected after data collection?
Rarely. Kesmodel (2018) states that few and insufficient methods exist for correcting inaccurate information, so prevention during planning matters most.
Sources
All sources checked 2026-10-29.
- Catalogue of Bias, Information bias.
- Catalogue of Bias, Misclassification bias.
- Catalogue of Bias, Recall bias.
- Catalogue of Bias, Observer bias and Reporting bias.
- Kesmodel US. Information bias in epidemiological studies with a special focus on obstetrics and gynecology. Acta Obstet Gynecol Scand. 2018;97(4):417-423. doi:10.1111/aogs.13330; abstract at PubMed 29453880.
- Chang et al. (2010), daily self-reports and exposure duration to computer use, as summarised on the Catalogue of Bias information bias page.
