Information bias
How it differs from selection bias
Selection bias concerns who ended up in the study. Information bias concerns what was recorded about the people who did. They can both be present and they need different remedies, so naming which one a limitation refers to is worth doing.
The recurring forms
- RECALL bias: cases search their memory harder than controls, systematically in case-control studies of past exposures.
- OBSERVER bias: an assessor who knows the group assigns outcomes differently, especially for subjective endpoints.
- DETECTION or SURVEILLANCE bias: one group is looked at more, so more is found. A drug requiring monitoring blood tests will appear associated with anything those tests detect.
- INTERVIEWER bias: knowledge of case status shapes how questions are asked and followed up.
- REPORTING and social desirability bias: alcohol, adherence and diet are all reported in the direction the respondent thinks is expected.
Design defences
Blind outcome assessors wherever an outcome involves judgement, and say so; use objective or independently adjudicated endpoints; collect exposure data before the outcome is known, which is why prospective and nested designs are preferred; apply identical ascertainment intensity to both groups by protocol; and use a comparator with a similar monitoring schedule so surveillance is matched rather than confounded with exposure.
Quantify rather than mention
Where the bias cannot be designed away, a quantitative bias analysis over plausible values of sensitivity and specificity in each group shows how large the effect would have to be to explain the finding. That is more useful than a limitations paragraph, and it is often reassuring.
What to watch out for
The dangerous forms are the ones where the error depends on the group, because those bias in a direction the study cannot detect from its own data. Surveillance bias is the most easily overlooked: patients under closer follow-up accumulate more diagnoses, so an exposure that brings people into contact with the health system manufactures an association with anything that gets found by looking. Blinding the outcome assessor is cheap and is left out of most observational protocols.
Clinical example
A cohort finds a drug associated with more diagnosed atrial fibrillation. The drug requires quarterly review with an ECG; the comparator does not. The additional diagnoses may be entirely a product of looking more often, and the design that answers this matches the monitoring schedule between groups.
- Effects of non-differential exposure misclassification on false conclusions in hypothesis-generating studies (opens in a new tab) · Int J Environ Res Public Health · doi 10.3390/ijerph111010951
- Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD statement (opens in a new tab) · BMJ · doi 10.1136/bmj.g7594