Definition A systematic error in collecting or interpreting observations found in the study design Types of Bias Types of BiasBiasDescriptionMitigationAccumulation Effect<ul style="font-size: medium;"><li>patients sometimes must be exposed to a risk factor for aprolonged period of time before they develop a clinically detectable result<ul><li>e.g., patients must smoke for many pack-years before bronchogenic carcinoma develops</li></ul></li></ul><ul><li>try to follow study participants for as long as is feasible</li></ul>Confounding<a style="font-size: medium;" title="question" href="#210433" link-extracted="true"><img src="/images/question.png" alt=""></a><ul style="font-size: medium;"><li>a third factor is either positively or negatively associated with both the exposure and outcome</li><li>confounders arenot in the causal pathway<ul><li>if not adjusted for, can distort true association</li><li>either towards or away from the null hypothesis</li></ul></li></ul><ul style="margin: 16px 0px; padding: 0px 0px 0px 20px; border: 0px; vertical-align: baseline; list-style-position: initial; list-style-image: initial; font-size: small;"><li style="margin: 0px; padding: 0px; border: 0px; vertical-align: baseline;">randomization<a style="font-size: medium;" title="question" href="#210434" link-extracted="true"><img src="/images/question.png" alt=""></a><ul style="font-family: Arial, Helvetica-Neue, Helvetica, sans-serif; font-size: small;"><li style="margin: 0px; padding: 0px; border: 0px; vertical-align: baseline;">ensures similar baseline characteristics between control and exposure/experimental groups</li><li style="margin: 0px; padding: 0px; border: 0px; vertical-align: baseline;">use intention-to-treat analysis to preserve randomization even if participants change study treatments<a style="font-size: medium;" title="question" href="#210435" link-extracted="true"><img src="/images/question.png" alt=""></a></li></ul></li><li style="margin: 0px; padding: 0px; border: 0px; vertical-align: baseline;">matching<ul><li style="margin: 0px; padding: 0px; border: 0px; vertical-align: baseline;">group similar participants into study pairs</li></ul></li><li style="margin: 0px; padding: 0px; border: 0px; vertical-align: baseline;">stratification<ul><li style="margin: 0px; padding: 0px; border: 0px; vertical-align: baseline;">analyze in separate subgroups determined by a potential confounder</li></ul></li><li style="margin: 0px; padding: 0px; border: 0px; vertical-align: baseline;">restriction<ul><li style="margin: 0px; padding: 0px; border: 0px; vertical-align: baseline;">only include groups with specific features in the sample</li></ul></li><li style="margin: 0px; padding: 0px; border: 0px; vertical-align: baseline;">adjustment<ul style="font-size: small;"><li style="margin: 0px; padding: 0px; border: 0px; vertical-align: baseline;">can only adjust for confounders that areknown and measureable</li></ul></li><li>crossover studies<ul><li>subject acts as own control</li></ul></li></ul>Selection Bias <ul style="font-size: medium;"><li style="font-size: medium;">sampled population is not representative of the population researchers are trying to study<ul><li style="font-size: medium;">due to non-random selection of study participants</li><li>sampling (ascertainment) bias<ul><li>certain individuals are more or less likely to be selected for a study group, leading to incorrect conclusions</li><li>non-response bias<ul><li>e.g., participants who pick up the phone may be less sick than participants who don't</li></ul></li><li>healthy worker effect<ul><li>samples with employed subjects only may be healthier</li></ul></li><li>volunteer bias<ul><li>people who volunteer for a study may be different in some fundamental way from those who do not volunteer</li></ul></li></ul></li><li>late-look bias<ul><li>patients with severe disease are less likely to be studied, because they die or are otherwise unavailable, making a disease look less severe<ul style="font-family: Arial, Helvetica-Neue, Helvetica, sans-serif; font-size: medium;"><li>e.g., a group of HIV+ individuals are all asymptomatic</li></ul></li><li>also can have opposite effect<ul><li>e.g.,people with more mild disease are cured before the study takes place and only persistently sick folks are included in the study, making a disease seem more severe</li></ul></li></ul></li><li>Berkson bias<ul><li>hospitalized study subjects are more likely to have a greater burden of illness than other possible subjects</li></ul></li><li>attrition bias<a title="question" href="#216600" link-extracted="true"><img src="/images/question.png" alt=""></a><ul><li>those lost to follow-up may be different from those who remain in the study</li></ul><ul style="font-size: medium; font-family: Arial, Helvetica-Neue, Helvetica, sans-serif;"></ul></li></ul></li></ul><ul style="margin: 16px 0px; padding: 0px 0px 0px 20px; border: 0px; vertical-align: baseline; list-style-position: initial; list-style-image: initial; font-size: small;"><li style="margin: 0px; padding: 0px; border: 0px; vertical-align: baseline;">randomization</li><li style="margin: 0px; padding: 0px; border: 0px; vertical-align: baseline;">include patients in multiple settings (outpatient, hospitalized)</li><li style="margin: 0px; padding: 0px; border: 0px; vertical-align: baseline;">study designs that are longitudinal in nature rather than cross-sectional</li><li style="margin: 0px; padding: 0px; border: 0px; vertical-align: baseline;">gather maximal information on participants</li></ul><p></p><p></p><p></p><p>Measurement Bias</p><ul style="font-size: medium;"><li>information is gathered in a way that distorts the information or misclassifies study participants<ul><li>interviewer bias<ul><li>subjects in one group are interviewed in a different way than another<ul><li>differences due to interviewing style disrepencies are falsely attributed to group differences</li></ul></li></ul></li></ul></li></ul><p></p><p></p><p></p><ul style="font-family: Arial, Helvetica-Neue, Helvetica, sans-serif; margin: 16px 0px; padding: 0px 0px 0px 20px; border: 0px; vertical-align: baseline; list-style-position: initial; list-style-image: initial; font-size: small;"><li style="margin: 0px; padding: 0px; border: 0px; vertical-align: baseline;">standardize data collection</li></ul>Recall Bias<a style="font-size: medium;" title="question" href="#215233" link-extracted="true"><img src="/images/question.png" alt=""></a><ul style="font-size: medium;"><li>subjects with the disease are more likely to recall the exposure of interest<ul><li>e.g., parents of children with cancer recall exposure to a chemical<ul></ul></li></ul></li></ul><ul style="font-family: Arial, Helvetica-Neue, Helvetica, sans-serif; margin: 16px 0px; padding: 0px 0px 0px 20px; border: 0px; vertical-align: baseline; list-style-position: initial; list-style-image: initial; font-size: small;"><li style="margin: 0px; padding: 0px; border: 0px; vertical-align: baseline;"> reducing follow-up time in retrospective studies</li></ul>Performance Bias<ul style="font-size: medium;"><li>researchers treat groups differently or subjects alter their behavior/responses due to study group awareness<ul><li>Hawthorne effect<a style="font-size: medium;" title="question" href="#215231" link-extracted="true"><img src="/images/question.png" alt=""></a><ul style="font-size: medium;"><li>subjects alter their behavior when they know they are being studied</li></ul></li><li>procedure bias<ul style="font-size: medium;"><li>researcher decides assigment of treatment versus control and assigns particular patients to one group or the other nonrandomly</li><li>patient decides assigment of treatment versus control</li></ul></li></ul></li></ul><ul style="font-family: Arial, Helvetica-Neue, Helvetica, sans-serif; margin: 16px 0px; padding: 0px 0px 0px 20px; border: 0px; vertical-align: baseline; list-style-position: initial; list-style-image: initial; font-size: small;"><li style="margin: 0px; padding: 0px; border: 0px; vertical-align: baseline;">blinding</li></ul><p>Lead-Time Bias</p><ul style="font-size: medium;"><li>subjects appear to survive longer when in reality their disease was detected earlier<ul><li>common with improved screening</li></ul></li><li>e.g.,a cancer screening test is deemed to increase survival when in reality the disease was picked up earlier, increasing the time from detection to death</li></ul><ul style="font-family: Arial, Helvetica-Neue, Helvetica, sans-serif; margin: 16px 0px; padding: 0px 0px 0px 20px; border: 0px; vertical-align: baseline; list-style-position: initial; list-style-image: initial; font-size: small;"><li style="margin: 0px; padding: 0px; border: 0px; vertical-align: baseline;">use mortality rate instead of survival time in screening studies</li><li style="margin: 0px; padding: 0px; border: 0px; vertical-align: baseline;">estimate lead time and add that to survival in unscreened group</li></ul>Design Bias<ul style="font-size: medium;"><li style="font-size: medium;">the control group is inappropriately non-comparable to the intervention group<ul><li>allocation bias<ul><li>difference in the way participants are placed in control versus experimental groups</li><li>e.g.,all zebras in control group and all lions in exposure group</li></ul></li></ul></li></ul><p></p><p></p><ul style="font-family: Arial, Helvetica-Neue, Helvetica, sans-serif; margin: 16px 0px; padding: 0px 0px 0px 20px; border: 0px; vertical-align: baseline; list-style-position: initial; list-style-image: initial; font-size: small;"><li style="margin: 0px; padding: 0px; border: 0px; vertical-align: baseline;">randomization</li><li style="margin: 0px; padding: 0px; border: 0px; vertical-align: baseline;">matching</li></ul>Cognitive Bias<ul><li style="font-size: medium;">observer bias (pygmalion effect)<a style="font-size: medium;" title="question" href="#215238" link-extracted="true"><img src="/images/question.png" alt=""></a><ul><li style="font-size: medium;">investigator inadvertently conveys her high expectations to subjects, who then produce the expected result<ul style="font-size: medium;"><li>a "self-fulfilling prophecy"</li><li>golem effect is the opposite: study subjects decrease their performance to meet low expectations of investigator</li></ul></li></ul></li><li>confirmation bias<ul><li>researcher ignores results that do not support their hypothesis</li></ul></li><li>response bias<ul><li>participants do not respond accurately because they are concerned about the social desirability of their responses or misinterpret the question</li></ul></li></ul><ul style="font-size: medium;"></ul><p></p><p></p><p></p><p></p><ul style="font-family: Arial, Helvetica-Neue, Helvetica, sans-serif; margin: 16px 0px; padding: 0px 0px 0px 20px; border: 0px; vertical-align: baseline; list-style-position: initial; list-style-image: initial; font-size: small;"><li> double blinding</li><li style="margin: 0px; padding: 0px; border: 0px; vertical-align: baseline;">include positive and negative results</li></ul>Surveillance Bias<ul><li>outcomes are more likely to be detected in certain groups because of increased monitoring<ul><li>e.g.,a certain skin disease being detected more often in hypertensive patients because they have more physician visits than non-hypertensive patients</li><li>researchers may falsely attribute hypertension to causing the skin disease</li></ul></li></ul><ul style="font-family: Arial, Helvetica-Neue, Helvetica, sans-serif; margin: 16px 0px; padding: 0px 0px 0px 20px; border: 0px; vertical-align: baseline; list-style-position: initial; list-style-image: initial; font-size: small;"><li style="margin: 0px; padding: 0px; border: 0px; vertical-align: baseline;">match participants on similar likelihood of surveillance</li></ul> Examples of Effects that are NOT Bias