| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Reviews |
1 From Harvard Vanguard Medical Associates, Boston, Mass. Received June 23, 2004; revision requested August 11; revision received August 22; final version accepted January 21, 2005; updated August 29. Address correspondence to the author, Department of Radiology, Woodhull Medical Center, 760 Broadway, Brooklyn, NY 11026 (e-mail: gsica{at}sprynet.com).
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONDESCRIPTIVE AND ANALYTIC STUDY...SELECTION BIASINFORMATION OR OBSERVATION BIASCOMMON METHODS FOR REDUCTION...CONCLUSIONESSENTIALSReferences |
|---|
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONDESCRIPTIVE AND ANALYTIC STUDY...SELECTION BIASINFORMATION OR OBSERVATION BIASCOMMON METHODS FOR REDUCTION...CONCLUSIONESSENTIALSReferences |
|---|
Readers must also be aware of bias—its effect on validity and how it can lead to data misinterpretation and limit the applicability or generalizability of a given study. The importance of a reader's ability to assess for bias in research studies was addressed in the recently reported guidelines of the Standards for Reporting of Diagnostic Accuracy, or STARD, initiative (2).
This review will (a) describe various types of bias that may be found in both the radiology and the general medical literature, (b) discuss bias in relation to specific study designs, and (c) report some common general methods to reduce bias. Much of the terminology used is derived from the epidemiology literature and is in common use within other medical specialties. While this may not be entirely familiar to radiologists, an effort was made to "bridge this gap." The overall goal is to assist readers in recognizing and assessing the magnitude and impact of bias on study results.
A common reason to undertake a scientific investigation is the intent to demonstrate the presence or lack of an association (eg, between an exposure and disease) or differences in parameter estimates (eg, means, standard deviations, proportions) between populations. In this setting, alternative explanations also must be excluded. Associations or differences may be real, may occur by chance (sampling error), or may be related to other factors such as bias (3).
Two broad types of error can affect scientific investigations and distort measurements: random and systematic. Because it is not feasible to study an entire population, a sample of the population is chosen. The study sample, however, may not accurately reflect the full spectrum of characteristics found in the target population. Random sampling error can then result and reflects variability or chance variation that may occur from sample to sample (4,5). Studies with a small sample size (eg, enrolled participants or number of imaging studies) are more prone to this type of error. For instance, if one were interested in studying prostate volume as measured with magnetic resonance (MR) imaging in men aged 70 years versus those aged 50, there would be less variability in the mean estimates if random samples of 1000 subjects were studied than if samples of 10 were studied.
Proper study design may help mitigate the effects of sampling error. Statistical tests aid in predicting and quantifying the magnitude of sampling variation and are used to assess associations or differences that occur by chance (3). Inference—the process of developing generalizations from sample data—can then be extended from the study sample population to the reference population.
Bias is a form of systematic error, and there are innumerable causes. The causes of bias can be related to the manner in which study subjects are chosen, the method in which study variables are collected or measured, the attitudes or preferences of an investigator, and the lack of control of confounding variables (a distortion of observed associations by additional, sometimes not readily apparent, variables). In epidemiologic terms bias can lead to incorrect estimates of association, or, more simply, the observed study results will tend to be in error and different from the true results. Bias should be considered primarily a function of the study process (ie, design and methods) and not of the results (5,6).
An unbiased study is considered to be valid—that is, the study results are, on average, correct. The distinction between bias and variation is not always obvious. Rothman (4) provides a conceptual model to distinguish systematic from random errors. Consider the hypothetic case in which a study sample or case size could be increased until it was infinitely large; chance variation of the mean, or random error, would be reduced toward zero. These are random errors. Systematic errors would not be diminished by increasing sample size.
|
DESCRIPTIVE AND ANALYTIC STUDY DESIGNS |
|---|
|
TOPABSTRACTINTRODUCTIONDESCRIPTIVE AND ANALYTIC STUDY...SELECTION BIASINFORMATION OR OBSERVATION BIASCOMMON METHODS FOR REDUCTION...CONCLUSIONESSENTIALSReferences |
|---|
Study designs can be broadly categorized as descriptive or analytic (Table 1). Descriptive studies typically describe general characteristics such as the imaging appearance of a disease in relation to various variables such as person, population, place, signs and symptoms, and disease severity and duration. Descriptive studies are often, but not always, retrospective in nature, as both disease and exposure are known at the time of the study. The data are usually readily available, and the study design is efficient. Retrospective descriptive studies are useful for investigations of rare diseases or new technologies. They tend not to raise ethical concerns because they do not require active patient participation and consent, do not result in the performance of additional imaging examinations, and do not increase patient risk (biopsies, laboratory work) and patient costs.
|
Analytic studies can be of an observational or intervention type (experimental or clinical trials). With observational studies, the investigator observes characteristics of exposure and outcome within or between groups; with intervention studies, the investigator allocates subjects to exposure or treatment groups and then follows them up for the development of an outcome. Outcomes of interest may include the development of a certain disease, the efficacy of a treatment, or the diagnostic accuracy of an imaging examination. The analytic category includes both retrospective designs (case-control and cohort studies) and prospective designs (cohort and intervention studies). The latter are typically considered to provide the highest quality data, but a well-designed case-control study can also achieve a high standard of validity (4). A general feature of analytic studies is that the use of an appropriate comparison group allows testing of hypotheses (10).
In case-control studies the investigator chooses cases and controls on the basis of the presence of a particular disease. Since both exposure and disease have occurred at the time of the study, subject selection or sampling biases are a particular concern. These types of studies are not uncommon in radiology and typically compare a certain imaging feature in diseased versus healthy populations. Patients can also serve as their own control by undergoing two different diagnostic tests. Some examples include diagnostic accuracy efficacy studies such as a comparison of MR imaging signal intensity in the prostate in healthy subjects versus in those with carcinoma or MR signal intensity and enhancement in the pancreas in healthy subjects versus in those with pancreatitis. Both exposure (inciting factors leading to disease or, in these imaging studies, the MR examination) and disease (carcinoma or pancreatitis) have already occurred.
With cohort studies, the investigator chooses subjects on the basis of exposure to a risk factor or intervention, rather than presence of disease, and exposed and unexposed individuals are followed up over a period of time to measure disease occurrence. Cohort studies have been less common in radiology research. For example, one might study the association between x-ray exposure among interventional radiologists and subsequent development of cancer in comparison with an unexposed cohort (eg, sonologists). If such a study such were prospective, it would take many years for the development of the outcome of interest (ie, cancer). Clinical trials and intervention studies can play an important role in helping establish higher level efficacy for new image-guided interventional therapeutic techniques, where patient outcomes are compared with those associated with conventional therapies (eg, intravascular stent repair vs conventional surgical repair for abdominal aortic aneurysm, radiofrequency ablation vs standard surgical resection for liver metastases). The prospective design of cohort and intervention studies is appealing in that it can minimize selection biases. These studies, though, tend to require larger number of subjects and longer observation times and may be more costly. In addition, loss to follow-up may be a particular source of bias.
The concept of efficacy has evolved over the past few decades. A six-tiered hierarchic model (Table 2) was initially described by Fryback (11), was comprehensively reviewed more recently (12), and is now commonly accepted. In general, efficacy in this model refers to the probability of benefit, to a specific population, of a given medical technology but is specifically defined within each tier. The efficacy approach expands on the more traditional goals of diagnostic imaging, which are to provide the highest technical quality and diagnostic accuracy (13). A key feature is that for an imaging examination to be efficacious at one level, it must be so at all lower levels (13). Other models exist (Table 1): for example, ranking of studies according to grades or quality of evidence (14). Concato et al (15) challenged the notion of a rigid hierarchic ranking and reported that well-designed and well-executed observational studies often can provide information at the level of quality of a randomized controlled trial.
|
The goal in developing a study design is not necessarily to eliminate all types of bias. In doing so, the effect may be to limit generalizability and render a study less useful. Further, if publication standards are set too high, "imperfect" studies that may contain useful or provocative information might not be available for scrutiny. More than 30 types or named variations of bias have been described (Figure), often with overlap in meaning (18). Terminology is not fully standardized and can be medical-specialty specific. While different types of bias overlap, it is useful to categorize them into broad related groups.
|
|
SELECTION BIAS |
|---|
|
TOPABSTRACTINTRODUCTIONDESCRIPTIVE AND ANALYTIC STUDY...SELECTION BIASINFORMATION OR OBSERVATION BIASCOMMON METHODS FOR REDUCTION...CONCLUSIONESSENTIALSReferences |
|---|
Eng and Siegelman (20) and Peterman (21) have described a multistep process that begins with defining the target population: that population for whom the study examination is intended. Inclusion and exclusion criteria then define the accessible population. A sampling scheme is devised and applied, which results in the intended sample population. Those who are ultimately enrolled or chosen yield the study population. In general, selection biases can be minimized in prospective studies and are more problematic in retrospective studies, because both disease outcome and exposure have already been ascertained at the time of participant selection (6).
Sample Bias
Sample bias can arise when the intended sample does not adequately reflect the spectrum of characteristics in the target population. The problem of sample bias also extends to the choice of control group in a case-control study. The control group should not necessarily consist of healthy disease-free volunteers but of those volunteers who are considered to be in the target population. Not doing so may result in overestimation of diagnostic accuracy or in the test in question appearing more specific than under less ideal clinical conditions (22,23).
Loss-to–Follow-up Bias
Loss-to–follow-up bias can be seen in cohort studies when subjects who are lost to follow-up differ from those who remain in the study until an event occurs or the study is terminated. The probability of the outcome of interest may differ in subjects lost to follow-up versus in those who remain in the study (5). In a prospective study with CT to determine the incidence of lung cancer in a high-risk exposed (smoker) population versus in a nonexposed (nonsmoker) population, multiyear follow-up is required. One could envision that after several years of negative studies, control subjects may be less motivated to continue in the study. Alternatively, exposed individuals may be at greater risk of developing comorbid disease and may discontinue study participation. It would be difficult to predict the differential magnitude of these losses prospectively. The analogous situation in a case-control study is when follow-up differences exist between the cases and controls.
Disease Spectrum Bias
Disease spectrum bias may occur when only cases within a limited range of a disease spectrum are included. This more commonly occurs with more obvious or advanced disease. Mild disease may be difficult to diagnose and may go undetected and inadvertently omitted. For example, in a study investigating the ability of MR imaging to depict cirrhosis, if only advanced clinical cases are included the sensitivity will be overestimated. Further, borderline-positive clinical cases that are MR negative may not undergo a definitive reference test (biopsy). In a case-control study, the comparison of very sick with healthy subjects may lead to exaggerated results. In that situation, the results will tend to overestimate the performance of the examination in question as applied to the target population. Such findings, though biased, might still be useful in the preliminary evaluation of a new test in helping to decide whether to proceed with higher level studies (24).
Referral Bias
Referral bias occurs when individual preferences or local practices determine which subjects undergo a certain imaging study or if only a subset of the general case population is referred, as might occur at a tertiary-care academic center. While referral bias can affect disease spectrum, there may be more subtle differences between the study population and the target population. This situation can also arise when distinct clinical groups use different indices of suspicion or indications when ordering a certain study. For example, if one were to study the role of pulmonary CT angiography in the evaluation of patients with shortness of breath, there may be differences between patients (eg, in duration and severity of symptoms) referred from the emergency room versus those referred from primary care physicians. Neurologists and primary care physicians may also differ in referral patterns for head CT in patients with headaches. Certain physician specialists may also adopt the use of a new imaging study (eg, coronary CT angiography) prior to general acceptance. New imaging studies that have not been universally accepted into clinical practice are particularly prone to this type of bias.
Participation Bias
Participation bias may result from factors that affect final enrollment of the intended sample. Not all subjects may agree to participate, or certain imaging studies or medical records may not be available for review. Further, investigator or study-design factors may influence participation. Participation bias could result when patients are inadvertently excluded because of personal time constraints (eg, imaging is only offered during select times). In the case of a retrospective review, if only cases with pathologic proof of a disease are included, those cases from subjects with the disease and who underwent imaging but were unable to undergo biopsy or were treated elsewhere would not be included in a search of a pathology data base. Despite controlling for other types of bias, unidentified differences between the intended sample and the sample of patients who actually participate may still exist.
Image-based Selection Bias
Image-based selection bias can exist when the inclusion of subjects is dependent on their having undergone a certain imaging study. This is a common bias in the radiology literature, where study populations are often selected on the basis of availability for imaging studies. This situation may also arise when a subject undergoes imaging at a community hospital but is treated at a tertiary care center and the examination results are not available for review. The study population may differ from those with the same disease or exposure but who did not undergo the imaging study or for whom the study is not available. Therefore, the study population may not be truly representative of the target population. Another form of this selection bias is when the interpretation of a test relies to some extent on real-time viewing (eg, endoscopy, fluoroscopy, ultrasonography). In a retrospective study, only selected images are typically available for review (25).
Study Examination Bias
Study examination bias relates to the exclusion of technically limited or incomplete studies or the prospective inclusion of only patients who are deemed competent to produce a technically adequate examination (19). This bias will result in an overestimation of sensitivity, and specificity may increase because false-positive "artifacts" are decreased. For example, in a study to evaluate renal MR angiography versus conventional angiography, if only subjects are included who can adequately suspend respiration during MR imaging, the performance of MR angiography will be overestimated. As Begg and McNeil (26) assert, if a test is not repeatable, the analytic treatment should be similar to that for equivocal test results. Uninterpretable test results should not be discarded, and at the minimum their frequency should be reported (27).
Self-Selection Bias
Self-selection bias has been described for screening studies and exists when study subjects are self-selected for enrollment. Differences may exist between those who volunteer and those who refuse participation (28). Volunteers may be more health conscious or even healthier than the general population, and this may favorably affect the efficacy of the screening study. It is difficult to predict and quantify differences between a volunteer and target population, and randomization is considered to be an effective tool to address this type of bias.
|
INFORMATION OR OBSERVATION BIAS |
|---|
|
TOPABSTRACTINTRODUCTIONDESCRIPTIVE AND ANALYTIC STUDY...SELECTION BIASINFORMATION OR OBSERVATION BIASCOMMON METHODS FOR REDUCTION...CONCLUSIONESSENTIALSReferences |
|---|
Misclassification can occur in imaging studies when the reference test is inaccurate. If there is no relationship between the results of the study and those of the reference test (conditionally independent), this can lead to nondifferential misclassification and an underestimation of the performance of the study test. If the two tests are correlated, differential misclassification may occur, and study test characteristics may be under- or overestimated, depending on the nature of the dependence.
Recall Bias
Recall bias occurs when exposure information is differentially misclassified for subjects with and for those without disease (4). Recall bias can be particularly problematic in studies where subjects are interviewed to collect information, as might occur in case-control and retrospective cohort studies. One example is a study correlating symptoms at presentation with findings at pulmonary CT angiography. Those patients with positive results and who are subsequently treated may better recall or even exaggerate their presentation symptoms, while those with negative studies may underestimate them. Similarly, in patients presenting with chest pain who are evaluated for myocardial infarction, differential symptom recall may occur in those with a positive versus in those with a negative final diagnosis. The direction of differential recall cannot always be predicted or discerned, and the bias can lead to under- or overestimation of the association between exposure (eg, shortness of breath, chest pain) and disease (6).
Interviewer Bias
Interviewer bias may arise in studies in which subjects are interviewed (eg, survey-type studies) or medical records are reviewed by an investigator who is also involved in the interpretation of a test result or determination of disease classification. The investigator may inadvertently "coach" subjects or selectively review entire medical records. A common practice to reduce this bias is to use investigators or clinical coordinators to collect this information, who are not otherwise involved in disease or test status determination.
Verification Bias
Verification bias, also referred to as work-up bias, refers to potential differences in the manner in which disease status is determined. For instance, the decision to perform a reference test may be based on the results of the study test. If the reference test is only applied in the case of a positive study test result, then, systematically, more sick subjects than healthy subjects undergo the reference test. The potential for verification bias will be related to the degree of dependence on the study test results. Thus, a good study test, which helps detect more positive cases, may actually result in greater verification bias, assuming subjects with negative study test results do not undergo the reference test (26). Verification bias can also arise when the reference test is invasive, such as biopsy or surgery, or might be considered unethical to apply in subjects in whom there is no suspicion of disease. There are many examples of this in the radiology literature: for example, in imaging studies for evaluation of liver lesions. Patients with negative studies and those with studies with benign-appearing lesions do not typically undergo the invasive reference test—biopsy—and one could argue there are circumstances in which an alternate standard of proof—follow-up imaging—is not equivalent. Economic considerations can also constrain to whom the reference standard is applied (25). Verification bias can also occur when subjects with a positive test result undergo a more thorough search for disease (eg, search for prostate carcinoma in patients with borderline elevated prostate-specific antigen levels) than do those with a negative test result.
In retrospective reviews of diagnostic studies, Lijmer et al (22) and Whiting et al (29) reported the effect of verification bias to be large and dependent on the quality of the reference tests used. Sensitivity and negative predictive value are improved when only patients with positive study test results undergo the reference test (30). If a different and poorly performing reference test is used for negative studies, specificity may also be increased. Accuracy can be increased when different verification standards are used. The method in which disease status is ascertained should be carefully designed and outlined. This type of bias can be avoided if the reference test is always performed prior to the investigative test. This is often not practical or possible (eg, surgery before a CT diagnosis of appendicitis). For some study designs, verification bias can be minimized with appropriate blinding between the reference test and study test results. When clinical follow-up is used for verification, an adequate time interval is chosen to determine disease status conclusively, although not to the extent that some subjects are lost to follow-up.
A variation of verification bias exists when an investigator limits subjects to only those with definitive proof (eg, pathologic specimen) of disease status. This does not necessarily improve study design and may introduce even greater bias in subject or case selection (26).
Follow-up Bias
Follow-up bias, also referred to as medical surveillance bias, can occur when subjects differentially undergo follow-up of disease status. Screening studies (eg, mammography) are prone to this bias, for instance, when patients with positive study test results undergo more intensive follow-up. On the other hand, patients with negative results perhaps do not undergo a specific reference test, or an imperfect reference test is used, and are subject to this bias if not followed up as diligently as patients with positive test results. This may lead to misclassification bias and again underscores the need to specify adequate and appropriate case follow-up in the study design.
Response Bias
Response bias may exist when missing data are present nonrandomly for study subjects (eg, disease-free cases that were not investigated as thoroughly) who were otherwise still included in the analysis. The results may be biased by the subject group for whom more complete data are available. With patient preference or satisfaction studies, subjects with disease may be more motivated to respond accurately and completely to surveys. Their responses, thus, may be disproportionately represented.
Reviewer Bias
Reviewer bias may take many forms and overlaps with other types of bias described below. Reviewer bias can occur when the person collecting or reviewing data (either subjective or objective) is inappropriately blinded or is aware of a suspected diagnosis or results of a reference test. This type of bias can also occur when reference test results are interpreted with knowledge of the study test results. For example, a pathologist might be more inclined to choose a certain diagnosis if the imaging appearance supports it. Reviews of the imaging literature have reported this to be a common bias (31).
Diagnostic-Review Bias
Diagnostic-review bias occurs when the reference test results are not definitive and the study test results affect or influence how the final diagnosis is established (26). For example, this might occur in the case of a positive study test result and a borderline-positive reference test result. This type of bias could have existed in early studies of the accuracy of pulmonary CT angiography, with conventional pulmonary angiography as the reference test. There is a component of subjective interpretation with pulmonary angiograms, particularly for distal filling defects. It is possible that a positive finding on the CT study test could influence the final interpretation of the reference test (conventional pulmonary angiography). Similarly, the final interpretation of a borderline-positive pathologic examination result (the reference test) could be influenced by strongly positive imaging findings (the study test).
Test-Review Bias
Test-review bias may arise in retrospective studies where the study test is performed after the diagnosis has been established. If the study test is subjective, knowledge of the diagnosis may affect test interpretation (30). This is potentially common in studies of imaging technology assessment, in which select populations with known disease are studied. For example, in studies to evaluate the accuracy of MR imaging for the detection of tumor in patients with biopsy-proved breast or prostate carcinoma, if only subjects with known disease are imaged and the image reader is aware of the disease status, this may influence the interpretation of the imaging studies. In addition to any subjective effect on investigator interpretation, these studies are often limited by the lack of an adequate control group (eg, biopsy-negative breast masses or prostate nodules).
Incorporation Bias
Incorporation bias occurs when the results of the study test are incorporated as evidence for the final diagnosis (30). This may occur when the reference test cannot be performed because of ethical, economic, or logistic reasons. In a study of the use of MR imaging to characterize liver lesions, those subjects with benign lesions (eg, cysts, hemangiomas) would likely not undergo a reference test (eg, biopsy), and investigators might consider the "typical appearance" on MR images as supportive or even diagnostic of a benign condition. This type of bias can also occur when the reference test is not 100% accurate and the study test results are used as evidence for the final diagnosis (eg, in the case of CT angiogram that is positive for pulmonary embolus and a conventional pulmonary angiogram that is questionably positive). This type of bias can easily affect the apparent accuracy of the study test.
Imperfect-Standard Bias
Imperfect-standard bias occurs when the reference standard is not 100% accurate (eg, pulmonary angiogram for the diagnosis of pulmonary embolus) or a surrogate reference test is used in some subjects owing to cost or ethical considerations. An example is the measurement of colonic polyps by using CT colonography, with conventional colonoscopic measurements as the reference measurement. Since the colonoscopic measures are often subject to technical and human errors (eg, are estimated visually), they may not be accurate. Further, gastroenterologists vary in their estimates, so the bias is not applied uniformly. Obuchowski (27) described the example of the evaluation of multiple sclerosis with CT, with MR imaging as an imperfect reference test.
Occasionally, the results of a promising study test may even raise concern as to the accuracy of the accepted reference test (19). Interpretation errors of results from the imperfect standard may be independent of the study test (thus patients misclassified according to the study test have the same likelihood as all other patients as being misclassified according to the standard test), but when they are correlated (ie, they both result in misclassification of the same patients), the resulting sensitivity and specificity will be improved (32).
Reader-Order Bias
Reader-order bias can arise in comparative studies when retained knowledge of the results of one study influence the interpretation of the second study, potentially leading to a more accurate reading (eg, this might be seen when images obtained with multiple MR pulse sequences are sequentially compared for disease detection). Various randomization strategies can be used to address this bias; in addition, the introduction of a time lag between interpretations can further minimize its effect. A variant of reader-order bias is the potential for interpretation accuracy to improve if feedback is available to the reader during the course of a study owing to a learning-curve phenomenon.
Measurement Bias
Measurement bias reflects a discrepancy in measurements obtained with a new technique or test as compared with those obtained by using the reference technique or test with the same subjects under the same conditions and can occur with both objective and subjective measurements (16). Measurement bias can be uniform or nonuniform. A measuring tool, for instance, poorly calibrated electronic calipers, may introduce uniformly incorrect measurements.
With subjective assessments, nonuniform measurement bias may be introduced. Patient survey studies in which a nonvalidated questionnaire is used are particularly prone to this bias. If one wished to study the subjective perception of discomfort during conventional colonoscopy versus that during CT colonoscopy, the particular wording of questions may influence an individual subject's response (eg, there could be cultural or gender bias in the questions). When no reference standard exists, a scientifically sound method of validation must be devised that reflects the measurement in question.
Clustering Bias
Clustering bias, also referred to as repeated-measurement bias, occurs when multiple measurements or observations are obtained from the same subject. Such measurements can only be considered independent when obtained from individual subjects. When this is not the case, the data are considered to be correlated, and estimates of variability (eg, standard deviation) may be distorted. The extent of the bias relies on the magnitude of correlation and the number of observations (33).
While conventional statistics assume independence, there are statistical methods to adjust for a degree of clustering, some of which require certain assumptions (34). An example in a feature analysis-type imaging study is the inclusion of multiple liver lesions per subject. In a given subject, the liver lesions likely have similar pathobiologic characteristics and thus do not truly serve as independent measurements. Thus, one should either perform statistical tests to adjust for clustering or limit the number of lesions assessed per subject (which may result in a loss of power).
Context Bias
Context bias is related to the effect of altered disease prevalence on the estimates of given parameters such as sensitivity, specificity, and the receiver operating characteristics curve. For example, investigators are often required to review large numbers of selected imaging studies in a short time and in which disease prevalence is increased. The accepted belief is that sensitivity and specificity are not influenced by disease prevalence, thus allowing "transportability" of study results to other populations (32). While disease prevalence is known to affect positive and negative predictive values, Egglin and Feinstein (35) found that sensitivity and specificity may be affected, thus not merely changing position along a receiver operating characteristics curve but altering the curve itself. Other investigators have refuted this effect (36). This bias may particularly affect the interpretation of equivocal or difficult cases and persists even when spectrum and verification bias are taken into account.
Publication Bias
Publication bias may result if journals tend to publish studies with positive results or better-quality study designs. This bias does not arise within a given single study but can be seen in a review and analysis of the literature, as in a meta-analysis. This may lead to overly optimistic results or inflated associations (22).
Confounding
Confounding can be considered a type of bias; it results when additional factors or variables are associated with both exposure and, independently, with disease status, and these additional factors have a mixed effect. Confounding variables may or may not be obvious or known. They are a common problem in epidemiologic studies and particularly in observational studies where distribution of the variables among the exposed and unexposed subjects is unbalanced (37). Common confounding variables are age and sex. Confounding can be addressed in the study design by using techniques such as randomization, restriction, and matching, but if the effect is to limit the population evaluated, the study may become less generalizable.
Stratification is a common method to address confounding in the analysis phase (6). For example, an investigator might be interested in using MR imaging to study the relationship of myocardial mass (disease) as a function of age (exposure), with the hypothesis that myocardial mass decreases with age. Other variables such as gender and level of physical activity may also have an effect. Let us assume that male versus female sex and high versus low level of activity are associated with increased average mass. If the study groups are not stratified by these additional variables and are disproportionately distributed, the younger study group could be disproportionately composed of low-activity females; and the older study group, of active males. The overall affect would be to diminish the true association between age and myocardial mass. Confounding can be in either a positive or a negative direction, depending on the affect of a given variable.
|
COMMON METHODS FOR REDUCTION IN BIAS |
|---|
|
TOPABSTRACTINTRODUCTIONDESCRIPTIVE AND ANALYTIC STUDY...SELECTION BIASINFORMATION OR OBSERVATION BIASCOMMON METHODS FOR REDUCTION...CONCLUSIONESSENTIALSReferences |
|---|
Consecutive Recruitment
Nonconsecutive recruitment might inadvertently introduce characteristics that are not in proportion to those found in the target population. This may affect outcome measures and limit generalizability. In the example provided by Kazerooni (19), one can readily comprehend differences in patients who visit an emergency room for the evaluation of myocardial infarction symptoms in the morning versus those who visit during the night (eg, duration of symptoms may affect severity of disease).
Prospective versus Retrospective Studies
In a retrospective study, both disease or outcome and exposure have already been ascertained at the beginning of the study. Data collection cannot be retrospectively modified to minimize bias, and it is not always possible to discover fully the reasons subjects were referred for a certain imaging study or all relevant population characteristics. Retrospective studies, therefore, are prone to various biases, as discussed earlier, that can minimize the usefulness and impact of a given study.
With this is mind, it is important to define the study population and the methods of data collection and review for reader assessment. Prospective subject recruitment or data collection can be designed to reduce bias. An exception occurs when prospective data collection leads to unblinding of investigators and thus introduces other bias (28).
Blinding
Ensuring that information such as pertinent test results, demographic data, or disease status, which may affect an investigator's test interpretation or assessment of an outcome, is not available will minimize reader bias. A double-blinded study refers to one in which both the investigator and study subject are blinded to group assignment. In a review of diagnostic studies, Lijmer et al (22) found that inappropriate blinding led to an overestimation of diagnostic accuracy.
|
CONCLUSION |
|---|
|
TOPABSTRACTINTRODUCTIONDESCRIPTIVE AND ANALYTIC STUDY...SELECTION BIASINFORMATION OR OBSERVATION BIASCOMMON METHODS FOR REDUCTION...CONCLUSIONESSENTIALSReferences |
|---|
In this review, various types of bias and the terminology used were discussed. Both investigators and readers will benefit from a clear and concise presentation of study design, including methods of subject recruitment or image selection, review and analysis, and potential biases. While the reported hierarchies of study design imply a vertical stratification, each design serves an appropriate function and purpose, as long as the limitations of the study design are recognized.
|
ESSENTIALS |
|---|
|
TOPABSTRACTINTRODUCTIONDESCRIPTIVE AND ANALYTIC STUDY...SELECTION BIASINFORMATION OR OBSERVATION BIASCOMMON METHODS FOR REDUCTION...CONCLUSIONESSENTIALSReferences |
|---|
|
References |
|---|
|
TOPABSTRACTINTRODUCTIONDESCRIPTIVE AND ANALYTIC STUDY...SELECTION BIASINFORMATION OR OBSERVATION BIASCOMMON METHODS FOR REDUCTION...CONCLUSIONESSENTIALSReferences |
|---|
This article has been cited by other articles:
|
|
 |
|
  C. Keyzer, S. Pargov, D. Tack, V. Creteur, P. Bohy, V. De Maertelaer, and P. A. Gevenois Normal Appendix in Adults: Reproducibility of Detection with Unenhanced and Contrast-Enhanced MDCT Am. J. Roentgenol., August 1, 2008; 191(2): 507 - 514. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 G. L. Hixson Sr, R. E. Hendrick, E. D. Pisano, M. J. Yaffe, and C. A. Gatsonis A Limitation of ACRIN DMIST Radiology, August 1, 2008; 248(2): 702 - 703. [Full Text] [PDF]
|
|
|
|
|
|
S. W. Anderson, E. Rho, and J. A. Soto Detection of Biliary Duct Narrowing and Choledocholithiasis: Accuracy of Portal Venous Phase Multidetector CT Radiology, May 1, 2008; 247(2): 418 - 427. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 Y. Wu, A. P. Furnary, and G. L. Grunkemeier Using the National Death Index to Validate the Noninformative Censoring Assumption of Survival Estimation Ann. Thorac. Surg., April 1, 2008; 85(4): 1256 - 1260. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
G. A. Zamboni, J. B. Kruskal, C. M. Vollmer, J. Baptista, M. P. Callery, and V. D. Raptopoulos Pancreatic Adenocarcinoma: Value of Multidetector CT Angiography in Preoperative Evaluation Radiology, December 1, 2007; 245(3): 770 - 778. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
| RADIOLOGY | RADIOGRAPHICS | RSNA JOURNALS ONLINE |
