资源描述
Bias, Confounding, and Effect ModificationSTAT 6395Spring 2008 Filardo and Ng Any systematic error in the design or conduct of a study that results in a mistaken estimate of the association between an exposure and a diseaseBias is often a major problem in observational epidemiologic studiesBias Example: an association between an exposure an a disease in which the true relative risk is 2.0Systematic error (bias) is different than random error If the design and conduct of a study are unbiased, and there is no confounding, and we repeat the study an infinite number of times, the mean relative risk will be 2.0, with the individual relative risks from the different studies fluctuating around 2.0 Systematic error (bias) is different than random error If the design or conduct of the study is biased, and we repeat the study an infinite number of times, the mean relative risk will differ from 2.0 (for example, it may be 1.2), with the individual relative risks from the different studies fluctuating around 1.2Systematic error (bias) is different than random error Due to random variation, an association that is far from the truth can be observed in an unbiased study, but it usually wont be.Systematic error (bias) is different than random error Due to random variation, the true association can be observed in a biased study, but it usually wont be Systematic error (bias) is different than random error Statistical significance does not protect against biasSystematic error (bias) is different than random error Selection bias Information biasTwo major categories of bias Error that results from criteria or procedures used to select study subjects or from factors that influence study participation. With selection bias, the relation between exposure and disease is different for those who are selected for and participate in the study and those who should be theoretically eligible to participate.Selection bias Selection bias can occur as a result of: Incorrect selection criteria for study subjects Differences in characteristics between eligible subjects who agree to participate and eligible subjects who do not participateSelection bias Error due to collection of incorrect information about study subjects. Due to this incorrect information, subjects are classified into incorrect exposure or disease categories.Information bias Selection bias is a major issue in case-control studies Source population: the population that gives rise to the cases Selection bias is a major issue in case-control studies Cases should be selected such that the distribution of the exposures of interest among the cases selected for the study is the same as it is among all cases that arise in the source population. The cases should be representative of all cases that arise in the source population with respect to the exposures of interest. Controls should be selected such that the distribution of the exposures of interest among the controls is the same as it is in the source population. The controls should be representative of the source population with respect to the exposures of interest.Selection bias in case-control studies (cont.) Selection bias occurs when either: The cases are not representative of all cases that arise in the source population with respect to the exposures of interest and/or The controls are not representative of the source population with respect to the exposures of interest.Selection bias in case-control studies (cont.) In the hypothetical data depicted in the following tables, we will assume there is: no information bias, confounding, or random variabilitySelection bias in case-control studies: how it worksso that all differences are due to differences in selection of cases or controls Gold standard OR = 4.5Hypothetical case-control study including all cases and all non-cases from Source Population A Unbiased OR = (350 x4,500)/(500 x700) = 4.5This is an unbiased odds ratio because the selection of cases and controls was unrelated to exposure.Hypothetical case-control study including a 70% unbiased sample of the cases and 0.5% unbiased sample of the controls from Source Population A Biased OR = (350 x4,050)/(950 x700) = 2.13Selection of controls was related to exposure-over selecting exposed controls biases OR downwardSelection bias in choosing controls in a hypothetical case-control study including a 70% sample of the cases and 0.5% sample of the controls from Source Population A Example: A hospital-based case-control study of the relation of smoking to a given disease. Selection bias in choosing controls in a case-control study due to incorrect criteria for control selection If the control group includes persons hospitalized for smoking-related diseases (e.g, cardiovascular disease) the control group would likely have a higher proportion of smokers than the source population, and the resultant odds ratio would be biased downwardSelection bias in choosing controls in a case-control study due to incorrect criteria for control selection Selection bias in choosing controls in a case-control study due to a difference in participation rates between exposed controls and nonexposed controls Example: Case-control study of the relation between housing characteristics and lead poisoning among children 6 years of age or younger who are screened for blood lead levels at the Hill Health Center in New Haven Selection bias in choosing controls in a case-control study due to a difference in participation rates between exposed controls and nonexposed controls Cases: all children with a blood lead level of 10 micrograms/dL Controls: a systematic sample of children with a blood lead level of 10 micrograms/dL Housing characteristics and lead poisoning (cont.) Incentive for participation: the parents of the children were offered a free lead inspection of their homes Participation rate among cases: 91% (parents were motivated by their childs elevated blood lead level to have the inspection) Housing characteristics and lead poisoning (cont.) Participation rate among controls: 69% (parents did not have the same motivation to participate) The condition of the housing of the control parents who refused to participate was better than the condition of the housing of the control parents who did participate The housing of the controls selected for the study was in poorer condition than the housing of the source population The odds ratio for the association between measures of dilapidated housing and childhood lead poisoning would be biased downwardHousing characteristics and lead poisoning (cont.) Although the criteria for selecting controls were sound, the difference in participation rate between exposed controls and nonexposed controls resulted in a biased odds ratio Housing characteristics and lead poisoning (cont.) Selection bias in choosing cases in a hypothetical case-control study including a 70% sample of the cases and 0.5% sample of the non-cases from Source Population ABiased OR = (450 x4,500)/(500 x600) = 6.75 Selection of cases was related to exposure-over-selecting exposed cases biases OR upward Example: Population-based case-control study of pancreatic cancer cancer Hypothesis: vitamin C protects against development of pancreatic cancerVitamin C intake assessed by food frequency questionnaireSelection bias in choosing cases in a case-control study Median interval between diagnosis and interview: 9 months One-year case fatality rate of pancreatic cancer: 80%Many cases would die before being interviewedSelection bias in choosing cases in a case-control study Suppose vitamin C intake improves survival from pancreatic cancer Then vitamin C intake among cases selected for the study would be higher than vitamin C intake among all cases Over-selection of exposed cases would bias OR upwardSelection bias in choosing cases in a case-control study To avoid biased odds ratios, investigators often attempt to equalize selection bias between cases and controls by selecting cases and controls undergoing the same selection processesCompensating Selection Bias Compensating bias in choosing cases and controls in a hypothetical case-control study including a 70% sample of the cases and 0.5% sample of the non-cases from Source Population A Unbiased OR = (450 x4,286)/(714x600) = 4.5 Equal over-selection (1.5x) of exposed cases and controls Unbiased OR = (350 x4,500)/(500 x700) = 4.5This is the original table Hypothetical case-control study including a 70% unbiased sample of the cases and 0.5% unbiased sample of the controls from Source Population A Example: Cases and controls selected from among women attending a breast cancer screening programThese women are likely to have high prevalence of known breast cancer risk factors, (family history of breast cancer, history of benign breast disease, late age at first birth)Cases and controls undergoing the same selection processes in a case-control study of breast cancer Example: Cases and controls selected from among women attending a breast cancer screening programIf cases from this population were compared to controls from the general population, an overestimate of the magnitude of some risk factors would probably occurCases and controls undergoing the same selection processes in a case-control study of breast cancer Selecting both cases and controls from the screening program should make the bias the same in both groups, leading to unbiased odds ratiosThis is another way of saying that controls should be selected from the source population that gave rise to the cases Cases and controls undergoing the same selection processes in a case-control study of breast cancer Minimizing selection bias in case-control studies In the study design stage, carefully consider the criteria for selection of cases and controls, particularly with respect to ensuring internal validity Minimizing selection bias in case-control studies Choose study procedures aimed at maximizing the participation rate of the subjects selected for the study Selection bias would occur if participation were related to both exposure and the subsequent development of disease Because study participants are selected before the development of disease, this is unlikelyThe exposed group and nonexposed comparison group were drawn from the same source population and went through the same selection processSelection bias in cohort studies using internal comparison groups is unlikely The nurses who participated in the Nurses Health Study most likely differed from the nurses who did not, but since the same selection process was used to select the exposed group and the nonexposed internal comparison group, the relative risk estimates should be unbiased.Selection bias in cohort studies using internal comparison groups is unlikely Exposed cohort and nonexposed external comparison group are not selected from the same source populationThe exposed cohort may be selected such that it is at higher or lower risk for disease than the external comparison group for a reason other than the exposure of interestCohort studies using external comparison groups are prone to selection bias A selection bias in occupational cohort studies using a general population external comparison groupPersons selected for employment are usually healthier than and have lower mortality rates than the general population, which includes the sick and disabled.Healthy worker effect A selection bias in occupational cohort studies using a general population external comparison groupThe healthy worker effect makes any excess disease or mortality associated with an occupational exposure more difficult to detect than it would have been if a valid comparison group had been used, biasing the estimates of relative risk downwardHealthy worker effect When a subject in a cohort study is lost to follow-up, we do not know whether that subject developed the disease of interest during the remainder of the studys follow-up periodLosses to follow-up in cohort studies are analogous to selection bias in case-control studies If the subjects lost to follow-up have a different incidence of the disease of interest than the subjects not lost to follow-up, the estimates of the incidence rate of the disease of interest in the cohort will be biasedLosses to follow-up in cohort studies are analogous to selection bias in case-control studies However, relative risk estimates will be unbiased if the bias on the incidence rate estimates is the same in the exposed and nonexposed groups. A biased relative risk estimate will occur only if losses to follow-up are related to both disease and exposure The best defense against bias due to losses to follow-up is to make intense efforts to locate each cohort member, and thus minimize lossesLosses to follow-up in cohort studies are analogous to selection bias in case-control studies The best defense against bias due to losses to follow-up is to make intense efforts to locate each cohort member, and thus minimize lossesLosses to follow-up in cohort studies are analogous to selection bias in case-control studies Gold standard RR = 49.75/11.10 = 4.48Hypothetical cohort study with 100% follow-up (to keep the examples simple, we will not use the person-years method, but will use 10-year cumulative incidence) Unbiased RR = 49.75/11.10 = 4.48Hypothetical cohort study with 30% of the cohort lost to follow-up: losses to follow-up independent of exposure and disease Unbiased RR = 49.75/11.10 = 4.48Hypothetical cohort study with 40% of the exposed group and 20% of the nonexposed group lost to follow-up: losses to follow-up related to exposure, but not disease Unbiased RR = 37.36/8.33 = 4.48Hypothetical cohort study with 40% of those who developed disease and 20% of those who did not develop disease lost to follow-up: losses to follow-up related to disease, but not exposure Biased RR = 37.36/11.10 = 3.37Hypothetical cohort study: losses to follow-up related to disease and exposure Nondifferential exposure misclassification: misclassification of exposure unrelated to disease Nondifferential disease misclassification: misclassification of disease unrelated to exposure Differential misclassification: misclassification related to both exposure and diseaseInformation bias (error due to collection of incorrect information about study subjects) results in misclassification of exposure or disease Nondifferential misclassification tends to bias an association toward the null hypothesis (no association) Differential misclassification can bias an association either toward or away from the null hypothesis, depending on the specific nature of the misclassificationInformation bias (error due to collection of incorrect information about study subjects) results in misclassification of exposure or disease Inclusion of nonexposed subjects in the exposed group and exposed subjects in the nonexposed group will bias the relative risk toward the null if the exposure misclassificiation is unrelated to the future development of disease, which is usually the case Differential exposure misclassification is not likely in cohort studiesNondifferential exposure misclassification in a cohort study Gold standard RR = 49.75/11.10 = 4.48Hypothetical cohort study with 100% follow-up and 100% accuracy in exposure and disease classification Biased RR = 29.33/12.03 = 2.44Hypothetical cohort study with 20% of exposed misclassified as nonexposed and 10% of nonexposed misclassified as exposed, independent of disease: nondifferential exposure misclassification At baseline, study subjects complete a food frequency questionnaire about dietary habits over the past year.Measurement error due to imperfect recall will result in exposure misclassification which will occur in both the exposed and nonexposed groupNondifferential exposure misclassification in a cohort study: dietary assessment example Biased RR = 55.72/20.20 = 2.76Hypothetical cohort study with 0.1% of nondiseased misclassified as having developed the disease and 8% of the diseased misclassified as nondiseased, independent of exposure: nondifferential disease misclassification Biased RR = 99.50/15.09 = 6.59Hypothetical cohort study with 0.5% of nondiseased in the exposed group misclassified as having developed the disease and 0.04% of the nondiseased in the nonexposed group misclassified as having developed the disease: differential disease misclassification Disease misclassification is a particular issue when information on disease is obtained from the members of the cohort themselves (e.g. health questionnaire)Whenever possible, subject reports about disease should be confirmed by more objective means, such as review of medical recordsDisease misclassification in cohort studies Differential misclassification is a concern if the study members involved in data collection on disease or in disease classification are aware of the exposure status of the subjectsDisease misclassification in cohort studies Gold standard OR = 4.50Hypothetical case-control study with no misclassification of exposure or disease Biased OR = 3.54Hypothetical case-control study with 10% of cases misclassified as controls and 5% of controls misclassified as cases, independent of exposure: nondifferential disease misclassification Definitive diagnosis can only be made by brain biopsy, which isnt done. We therefore must rely for diagnosis on clinical criteria and exclusion of other diseases. The diagnostic criteria are imperfect and will result in misclassification of the disease statusNondifferential disease misclassification in case-control study: Alzheimers disease Persons with other types of dementia, such as multi-infarct dementia may be included in the case group. Persons with early Alzheimers disease may be included in the control groupNondifferential disease misclassification in case-control study: Alzheimers disease Biased OR = 5.31Hypothetical case-control study with 10% of exposed controls misclassified as cases and 1% of nonexposed controls misclassified as cases: differential disease misclassification Exposure: hypertensionHypertension is a risk factor for multi-infarct dementia, which could be confused with Alzheimers diseaseDifferential disease misclassification in case-control study: Alzheimers disease Classifying exposed persons as being nonexposed and nonexposed persons as being exposed will bias the odds ratio toward the null if the exposure misclassification is unrelated to disease status Classifying exposed persons as being nonexposed and nonexposed persons as being exposed can bias the odds ratio in either direction if the exposure misclassification depends on disease statusExposure misclassification in a case-control study: an important source of both nondifferential and differential misclassification Biased OR = 1.96Example: dietary assessmentHypothetical case-control study with 20% of the nonexposed misclassified as exposed and 16% of the exposed misclassified as nonexposed, independent of disease: nondifferential exposure misclassification Biased OR = 5.16Example: Recall biasHypothetical case-control study with 20% of the nonexposed cases misclassified as exposed and 5% of the nonexposed controls misclassified as exposed: differential exposure misclassification Recall bias Reporting bias Observer biasTypes of information bias that can lead to differential misclassification Systematic error due to differences in accuracy of recall of past exposures or diseases between study groups Example: family history of prostate cancer in a case-control study of prostate cancerRecall bias Men diagnosed with prostate cancer are often more aware of their family history than men who have not had prostate cancerIn a case-control study, reporting of family history of prostate cancer could be more complete among cases than among controls, biasing the result away from the null hypothesisRecall bias Systematic error due to selective revealing or suppression of information about exposure or disease due to attitudes, beliefs, or percep
展开阅读全文