Information bias in health research: definition, pitfalls, and adjustment methods

Information bias in health research: definition, pitfalls, and adjustment methods Alaa Althubaiti

Tag layer of the DHSS Meta-Tagging Project. The article text is reproduced from the staged copy named in sourceDesc and is not altered; the tags are held separately in standOff so that overlapping and nested spans are preserved.

Information bias in health research: definition, pitfalls, and adjustment methods 2016 Journal of Multidisciplinary Healthcare 10.2147/jmdh.s104807 althubaiti2016-information-bias-in-health-researc althubaiti2016-information-bias-in-health-researc.txt, the project's staged copy b01907f38e82550d4cc8aea23d181d9c10b21d97 cc-by-nc epidemiology

the paper's own words

the tag layer; a sentence written by the tagger about the paper, always anchored to a quotation but not itself a quotation

0 spans located by searching the stored text for the quote's exact words, allowing any whitespace between them A CONCEPT WHOSE OWN VOCABULARY IS THE EXAMPLE - information bias, itself 'otherwise known as misclassification', whose two main subtypes carry four names each in a single sentence apiece, with the alternatives listed and never distinguished Quotation withheld because this record is not cleared for redistribution. which forms of information bias arise in observational and experimental medical research, how each is recognised, and what can be done about it at design, measurement and analysis Quotation withheld because this record is not cleared for redistribution. a short taxonomy of information bias - self-reporting bias (social desirability, recall), measurement error bias, confirmation bias - each with its named alternatives, its typical study design, and an adjustment or prevention strategy Quotation withheld because this record is not cleared for redistribution. method Quotation withheld because this record is not cleared for redistribution. systematic error introduced by how a measurement is obtained rather than by whom it is obtained from - and which is often introduced unintentionally, so that it is hard to recognise even by the researcher who caused it Quotation withheld because this record is not cleared for redistribution. review review Quotation withheld because this record is not cleared for redistribution. ONE ERROR, FOUR NAMES, IN ONE SENTENCE: `Quotation withheld` Four names offered as interchangeable, with no test given for telling them apart - and one of them, `measurement bias`, names the consequence while the others name the cause Quotation withheld because this record is not cleared for redistribution. AND AGAIN FOR THE THIRD FORM: `Quotation withheld` And the paper's subject itself: `Information bias, otherwise known as misclassification`. Three concepts, four, four and two names, in one short review - the corpus holds papers that count 138 definitions of a term; this is the same phenomenon at the scale of a synonym list, and it is not remarked on Quotation withheld because this record is not cleared for redistribution. AND THE TOP-LEVEL DEFINITION IS A GENUS AND A DIFFERENTIA: `Quotation withheld In health studies, bias can arise from two different sources; the approach adopted for selecting subjects ... or the approach adopted for collecting or measuring data ... termed as selection bias and information bias.` The partition is by WHERE THE ERROR ENTERS, not by what it does to the estimate - which is why the same arithmetic consequence can belong to either Quotation withheld because this record is not cleared for redistribution. AND A DISTINCTION THE FIELD IS EXPLICITLY WARNED NOT TO COLLAPSE: `Quotation withheld` The warning exists because the two are routinely conflated, and the cost of conflating them is that the standard toolbox appears to cover both and covers one Quotation withheld because this record is not cleared for redistribution. AND A TERM DEFINED BY WHAT IT IGNORES: `An analysis that ignores the effect of measurement error on the results can be referred to as a naive analysis.` Its consequences are given as `inconsistent (or biased) and/or inefficient estimators of regression parameters`. A name for the default practice, which is the rhetorical move that makes an omission visible Quotation withheld because this record is not cleared for redistribution. AND THE MECHANISM OF RECALL BIAS IS SPELLED OUT AS A CHAIN: `in a case-control study, Quotation withheld As such, true exposure might be underreported in healthy controls and overreported in the cases. The size of the difference ... will consequently be inflated, and, in turn, the observed odds ratio would also increase.` The bias has a direction, and the direction is predictable from the design - which is what makes it adjustable at all Quotation withheld because this record is not cleared for redistribution. AND THE FIX FOR SELF-REPORTING IS VALIDATION AGAINST SOMETHING ELSE: `internal validation` compares responses `with other data collection methods, such as laboratory measurements`; `external validation` uses `medical record checks or reports from family or friends`. The instrument is checked against a second instrument, never against the truth - and the paper then says the laboratory measurements are themselves `often subject to measurement errors` Quotation withheld because this record is not cleared for redistribution. AND THE VALIDATION HAS ITS OWN SAMPLING PROBLEM: `It is important to have a random selection of individuals so that results from the validation can be generalized to any group of participants. When the sampling approach is not random and subjective, the results from the validation study can only apply to the same group of individuals.` A correction factor inherits the scope of the study that produced it - polman2011's MAGNIMS refusal, in a different discipline Quotation withheld because this record is not cleared for redistribution. AND A WHOLE DESIGN IS DEMOTED ON THE STRENGTH OF ONE BIAS: `bias often accompanies results from case-control studies. Hence, case-control studies can be conducted in order to generate a research hypothesis, but not to evaluate prognoses or treatment effects.` A study design assigned a permitted use - the same move donnelly2019 makes about a definition, here made about a method Quotation withheld because this record is not cleared for redistribution. AND THE ONE FIELD THAT DOES ADJUST FOR MEASUREMENT ERROR IS NAMED: `Quotation withheld` Where a number has to survive cross-examination, its error is modelled - the practice follows the consequence, not the methodology Quotation withheld because this record is not cleared for redistribution. AND THE AGREEMENT BETWEEN SELF-REPORT AND RECORDS IS DISEASE-SPECIFIC: `a high agreement was found between self-reported questionnaires and medical record diagnoses of diseases such as diabetes, hypertension, myocardial infarction, and stroke but not for heart failure.` The instrument is not uniformly valid across the things it measures, which is exactly what prigerson2009's differential item functioning analysis tests for and deletes Quotation withheld because this record is not cleared for redistribution. THE FOUR NAMES ARE LISTED AND NEVER SEPARATED. `measurement error, instrumental error, measurement imprecision, or measurement bias` are offered as one thing; `confirmatory, ascertainment, or observer bias` likewise. Some of these are not synonyms in ordinary use - *instrumental* error is a subclass of measurement error, and *observer* bias is usually distinguished from *confirmation* bias by who holds the belief. A review of error that hands the reader an unsorted synonym list is reproducing the problem it describes Quotation withheld because this record is not cleared for redistribution. AND THE ADVICE IS ADMITTED TO BE HARD TO FOLLOW: `Bias is often not accounted for in practice. Even though a number of adjustment and prevention methods ... are available, applying them can be rather challenging due to limited time and resources. For example, measurement error bias properties might be difficult to detect, particularly if there is a lack of information about the measuring instrument.` The adjustment methods all require knowing the error distribution, and the error distribution requires a validation study that is `expensive` Quotation withheld because this record is not cleared for redistribution. AND THE VALIDATION CHAIN BOTTOMS OUT ON ANOTHER MEASUREMENT: laboratory measurements are the recommended standard for validating self-report, and they are `often subject to measurement errors` - `in a validation study to estimate sodium intake using a 24-hour urinary excretion method, the estimated sodium intake tended to be lower than the true amount`. There is no step in the chain that touches the true value, and the paper says so about its own recommended remedy Quotation withheld because this record is not cleared for redistribution. AND CONFIRMATION BIAS IS THE ONE WITH NO ADJUSTMENT METHOD AT ALL: the remedies offered are independent checks, blinding, and `specific education and training programs`, and the conclusion is `it is important to accept that bias is a part of investigations. Quantifying this inevitable bias and its potential sources must be part of well-developed conclusions.` The paper's title promises adjustment methods; for the third of its three biases there are none Quotation withheld because this record is not cleared for redistribution. validating a self-reporting instrument against laboratory measurements (internal) or medical records and family report (external) before using it eliminates: social desirability bias in answers about drug use, income, diet or violence · outcome: partial Quotation withheld because this record is not cleared for redistribution. requiring the validation sample to be randomly selected eliminates: a correction factor that applies only to the group it was derived from · outcome: held Quotation withheld because this record is not cleared for redistribution. choosing a short recall period, stratifying it by how many events a participant had, and using memory aids and diaries eliminates: recall error scaling with the length of the period being remembered · outcome: partial Quotation withheld because this record is not cleared for redistribution. blinding or masking, single or double, and independent checks across laboratories eliminates: confirmation bias in any step that requires human judgment · outcome: partial Quotation withheld because this record is not cleared for redistribution. Bias Bias Quotation withheld because this record is not cleared for redistribution. any systematic error in the design, conduct, or analysis of a study any systematic error in the design, conduct, or analysis of a study Quotation withheld because this record is not cleared for redistribution. selection bias selection bias Quotation withheld because this record is not cleared for redistribution. information bias information bias Quotation withheld because this record is not cleared for redistribution. measurement error measurement error Quotation withheld because this record is not cleared for redistribution. measurement bias measurement bias Quotation withheld because this record is not cleared for redistribution. confirmation bias confirmation bias Quotation withheld because this record is not cleared for redistribution. observer bias observer bias Quotation withheld because this record is not cleared for redistribution. random sampling random sampling Quotation withheld because this record is not cleared for redistribution. measurement error variability measurement error variability Quotation withheld because this record is not cleared for redistribution. An analysis that ignores the effect of measurement error on the results An analysis that ignores the effect of measurement error on the results Quotation withheld because this record is not cleared for redistribution. a naïve analysis a naïve analysis Quotation withheld because this record is not cleared for redistribution. cases cases Quotation withheld because this record is not cleared for redistribution. healthy controls healthy controls Quotation withheld because this record is not cleared for redistribution. The size of the difference between the observed rates of exposure to risk factors in cases and controls The size of the difference between the observed rates of exposure to risk factors in cases and controls Quotation withheld because this record is not cleared for redistribution. the observed odds ratio the observed odds ratio Quotation withheld because this record is not cleared for redistribution. `bias` - `any systematic error in the design, conduct, or analysis of a study`; split by WHERE it enters into `selection bias` and `information bias` Quotation withheld because this record is not cleared for redistribution. `information bias` - `otherwise known as misclassification`; error originating in `the approach that is utilized to obtain or confirm study measurements` Quotation withheld because this record is not cleared for redistribution. `measurement error` = `instrumental error` = `measurement imprecision` = `measurement bias` - four names presented as interchangeable in one sentence, with no criterion for choosing between them Quotation withheld because this record is not cleared for redistribution. `confirmation bias` - `otherwise known as confirmatory, ascertainment, or observer bias`: `a decision is made according to the subject's preconceptions, beliefs, or preferences` Quotation withheld because this record is not cleared for redistribution. `naive analysis` - `An analysis that ignores the effect of measurement error on the results`. A name given to the default practice so that it can be argued against Quotation withheld because this record is not cleared for redistribution. `systematic` vs `random` error - systematic deviations are `either consistently higher or lower than the true values`; random ones are `not consistent`, follow a distribution, and are called `multiplicative or non-gaussian` when proportional to the measurement Quotation withheld because this record is not cleared for redistribution. extends → the existing bias literature, which the paper says has discussed types of bias while 'the problem of bias is still frequently ignored in practice' Quotation withheld because this record is not cleared for redistribution. critiques → the naive analysis - the standard practice of ignoring measurement error, which yields 'inconsistent (or biased) and/or inefficient estimators' Quotation withheld because this record is not cleared for redistribution. critiques → the use of case-control studies for anything beyond hypothesis generation, given that 'bias often accompanies results from case-control studies' Quotation withheld because this record is not cleared for redistribution. extends → forensic toxicology, named as the one field with 'the most theoretical understanding of measurement bias', from which medical research is asked to learn Quotation withheld because this record is not cleared for redistribution. measurement error —same-as→ measurement bias measurement error measurement bias the relation itself is separately evidenced · strength of the relation: 0.9 · note: 'If these errors…' Four names, and they are not obviously synonyms: instrumental error names a source, imprecision names a property of the instrument, and measurement bias names what the error does to the estimate. Tagged as the paper states it, and the un-sorted list is recorded as a limitation · the relation stated: measurement error, instrumental error, measurement imprecision and measurement bias are offered as names for one thing, in one sentence, with no test for telling them apart Quotation withheld because this record is not cleared for redistribution. confirmation bias —same-as→ observer bias confirmation bias observer bias the relation itself is separately evidenced · strength of the relation: 0.9 · note: 'Placing emphasis…' The second four-name list in one review. In ordinary use 'observer bias' is the observer's expectation affecting a measurement and 'confirmation bias' is a search strategy over evidence; here they are one entry · the relation stated: confirmation, confirmatory, ascertainment and observer bias are given as names for deciding according to one's preconceptions Quotation withheld because this record is not cleared for redistribution. Bias —operationalised-as→ any systematic error in the design, conduct, or analysis of a study Bias any systematic error in the design, conduct, or analysis of a study the relation itself is separately evidenced · strength of the relation: 0.95 · note: 'Bias can be defined…' Systematic is doing all the work - the definition excludes random error by construction, and the paper then spends its longest section on random measurement error, which is systematic in its effect on an estimate · the relation stated: bias is defined as systematic error anywhere in a study, which is what allows the split by point of entry rather than by effect Quotation withheld because this record is not cleared for redistribution. selection bias —distinct-from→ information bias selection bias information bias the relation itself is separately evidenced · strength of the relation: 0.95 · note: 'In health studies…' A clean two-way split with a stated criterion, which is more than most of the taxonomies in this corpus manage · the relation stated: the two sources are the approach to choosing subjects and the approach to measuring them - a partition by point of entry, not by consequence Quotation withheld because this record is not cleared for redistribution. random sampling —distinct-from→ measurement error variability random sampling measurement error variability the relation itself is separately evidenced · strength of the relation: 0.95 · note: 'However, RANDOM…' The warning is printed because the conflation is common, and the cost is specific: the confidence interval looks complete and covers one of two error sources · the relation stated: sampling variability is handled by standard statistical methods and measurement error uncertainty is not, so treating them as one hides the second Quotation withheld because this record is not cleared for redistribution. An analysis that ignores the effect of measurement error on the results —same-as→ a naïve analysis An analysis that ignores the effect of measurement error on the results a naïve analysis the relation itself is separately evidenced · strength of the relation: 0.9 · note: 'An analysis that…' Naming the default is how an omission becomes visible. Compare carroll2007's 'thinness' for missing fidelity data: both fields coined a word for a thing that is not there · the relation stated: the ordinary analysis, done without accounting for measurement error, is given a name - and the name is the argument Quotation withheld because this record is not cleared for redistribution. cases —contrasted-with→ healthy controls cases healthy controls the relation itself is separately evidenced · strength of the relation: 0.95 · note: 'study, CASES…' The bias is not noise; it has a sign that follows from the design, which is why the paper can say what it does to the odds ratio · the relation stated: cases recall exposure more readily than controls, so the same true history is reported at different rates on the two sides of the comparison Quotation withheld because this record is not cleared for redistribution. The size of the difference between the observed rates of exposure to risk factors in cases and controls —causes→ the observed odds ratio The size of the difference between the observed rates of exposure to risk factors in cases and controls the observed odds ratio the relation itself is separately evidenced · strength of the relation: 0.95 · note: 'The size of the difference…' A named error class traced all the way to the published statistic - three steps, each stated, which is what makes recall bias adjustable rather than merely deplorable · the relation stated: the recall asymmetry inflates the observed exposure difference, and the inflated difference raises the odds ratio the study reports Quotation withheld because this record is not cleared for redistribution. such errors are often called measurement error, instrumental error, measurement imprecision or measurement bias such errors are often called measurement error, instrumental error, measurement imprecision or measurement bias Quotation withheld because this record is not cleared for redistribution. confirmation bias is also called confirmatory, ascertainment or observer bias confirmation bias is also called confirmatory, ascertainment or observer bias Quotation withheld because this record is not cleared for redistribution. AND A DISTINCTION THE FIELD IS WARNED NOT TO COLLAPSE: `random sampling should not be confused with measurement error variability ... [standard methods] do not account for uncertainty due to measurement error` AND A DISTINCTION THE FIELD IS WARNED NOT TO COLLAPSE: `random sampling should not be confused with measurement error variability ... [standard methods] do not account for uncertainty due to measurement error` Quotation withheld because this record is not cleared for redistribution. AND A NAME FOR THE DEFAULT PRACTICE: `An analysis that ignores the effect of measurement error ... can be referred to as a naive analysis` AND A NAME FOR THE DEFAULT PRACTICE: `An analysis that ignores the effect of measurement error ... can be referred to as a naive analysis` Quotation withheld because this record is not cleared for redistribution. AND RECALL BIAS HAS A SIGN: cases `more likely to recall exposure ... than healthy controls`, so the difference `will consequently be inflated, and, in turn, the observed odds ratio would also increase` AND RECALL BIAS HAS A SIGN: cases `more likely to recall exposure ... than healthy controls`, so the difference `will consequently be inflated, and, in turn, the observed odds ratio would also increase` Quotation withheld because this record is not cleared for redistribution. AND ONE FIELD DOES ADJUST FOR IT: measurement error bias is `rarely ... adjusted for ... except in the field of forensic medicine`, where it is `particularly relevant for their type of research` AND ONE FIELD DOES ADJUST FOR IT: measurement error bias is `rarely ... adjusted for ... except in the field of forensic medicine`, where it is `particularly relevant for their type of research` Quotation withheld because this record is not cleared for redistribution. AND A STUDY DESIGN IS ASSIGNED A PERMITTED USE: `case-control studies can be conducted in order to generate a research hypothesis, but not to evaluate prognoses or treatment effects` AND A STUDY DESIGN IS ASSIGNED A PERMITTED USE: `case-control studies can be conducted in order to generate a research hypothesis, but not to evaluate prognoses or treatment effects` Quotation withheld because this record is not cleared for redistribution. AND SELF-REPORT AGREES WITH THE RECORD FOR SOME DISEASES AND NOT OTHERS: high agreement for `diabetes, hypertension, myocardial infarction, and stroke but not for heart failure` AND SELF-REPORT AGREES WITH THE RECORD FOR SOME DISEASES AND NOT OTHERS: high agreement for `diabetes, hypertension, myocardial infarction, and stroke but not for heart failure` Quotation withheld because this record is not cleared for redistribution. AND BIAS CAN BE DELIBERATE: `Often bias is unintentionally introduced into a study by researchers, making it difficult to recognize, but it can also be introduced intentionally. Thus, bias remains a very sensitive issue to address and discuss openly` AND BIAS CAN BE DELIBERATE: `Often bias is unintentionally introduced into a study by researchers, making it difficult to recognize, but it can also be introduced intentionally. Thus, bias remains a very sensitive issue to address and discuss openly` Quotation withheld because this record is not cleared for redistribution. `any systematic error in the design, conduct, or analysis of a study`, split into `selection bias` and `INFORMATION BIAS` by where the error enters bias Quotation withheld because this record is not cleared for redistribution. `otherwise known as misclassification` - it `originates from the approach that is utilized to obtain or confirm study measurements` information bias Quotation withheld because this record is not cleared for redistribution. four names in one sentence for `observed measurements [that] will differ from the actual values`, with no criterion offered for choosing between them measurement error Quotation withheld because this record is not cleared for redistribution. `otherwise known as confirmatory, ascertainment, or observer bias` - `a decision is made according to the subject's preconceptions, beliefs, or preferences` confirmation bias Quotation withheld because this record is not cleared for redistribution. `An analysis that ignores the effect of measurement error on the results` - a name for the default practice naïve analysis Quotation withheld because this record is not cleared for redistribution. `the often-marginalized measurement error bias` is the paper's own description of the second the forms of information bias covered Quotation withheld because this record is not cleared for redistribution. `The size of the difference ... will consequently be inflated` the direction of recall bias in a case-control study Quotation withheld because this record is not cleared for redistribution. self-report is validated against laboratory measurement; laboratory measurement is `often subject to measurement errors`; a 24-hour urinary excretion validation of sodium intake `tended to be lower than the true amount`. Every step compares one instrument with another, and the paper says so about its own recommended remedy a validation chain with no true value at the end of it Quotation withheld because this record is not cleared for redistribution. `Commonly used statistical methods can address the sampling variability during data analysis, but they do not account for uncertainty due to measurement error.` A confidence interval computed the usual way looks complete and covers one of two error sources what the standard toolbox silently does not cover Quotation withheld because this record is not cleared for redistribution. `the observed measurements deviate from true values in a consistent manner` - correctable by calibration against a reference measurement systematic error Quotation withheld because this record is not cleared for redistribution. `the deviation ... is not consistent`, follows a distribution, and when proportional to the measured amount is `referred to as multiplicative or non-gaussian errors` random error Quotation withheld because this record is not cleared for redistribution. the variable the reviewer can actually set - `a short recall period is preferable to a long one, particularly when asking participants about routine or frequent events` recall period Quotation withheld because this record is not cleared for redistribution.

The full text of this article is not redistributable, so it is not included here. The evidence sentences are withheld with it; the tag layer and its grounding status remain. The article itself is at https://doi.org/10.2147/jmdh.s104807.