Differential Misclassification Research Articles

Analysis of Undetectable HIV RNA Using Survival Analysis: Results Must Be Interpreted Carefully

This methodological note exemplifies the drawbacks of conventional survival analysis for studying the determinants of the evolution of HIV RNA after the initiation of a HAART regimen in observational cohort studies. The potential for loss of information might yield differential misclassification biases and therefore unreliable results. Longitudinal models are better suited to analyze the repeated measures of a continuous outcome such as HIV RNA.

Validation of self-reported screening mammography histories among women with and without breast cancer.

As part of a case-control study of the efficacy of screening mammography, the authors validated the mammography histories of 2,495 women aged 40-64 years with incident breast cancer diagnosed in 1994-1998 and a 25% random sample of 615 controls never diagnosed with breast cancer, all reporting a mammogram in the past 5 years. Subjects from five metropolitan areas of the United States were cross-classified by facility records ("gold standard") and self-report according to history of a recent screening mammogram (within 1 year or within 2 years). Sensitivity and specificity of self-reported screening at 1 year were 0.93 and 0.82, respectively, for cases and 0.92 and 0.80 for controls. At 2 years, sensitivity and specificity were 0.97 and 0.78 for both cases and controls. Confidence intervals for the differences in sensitivity and specificity were narrow and included zero. Scant evidence was found of telescoping (recollection of events as more recent than actual). Findings suggest that, in an interview-based case-control study of the efficacy of screening mammography, 1) estimated true prevalences of recent screening mammography adjusted for sensitivity and specificity will be slightly lower than self-reported prevalences, and 2) differential misclassification of exposure status is slight. Therefore, odds ratios will likely be biased toward the null, underestimating screening efficacy.

Misclassification in Logistic Regression with Discrete Covariates

AbstractWe study the effect of misclassification of a binary covariate on the parameters of a logistic regression model. In particular we consider 2 × 2 × 2 tables. We assume that a binary covariate is subject to misclassification that may depend on the observed outcome. This type of misclassification is known as (outcome dependent) differential misclassification. We examine the resulting asymptotic bias on the parameters of the model and derive formulas for the biases and their approximations as a function of the odds and misclassification probabilities. Conditions for unbiased estimation are also discussed. The implications are illustrated numerically using a case control study. For completeness we briefly examine the effect of covariate dependent misclassification of exposures and of outcomes.

Cotinine levels and self-reported smoking status in patients attending a bronchoscopy clinic

The reliability of self-reported smoking behaviour can vary and may result in bias if errors in misclassification vary with outcome. We examined whether self-report was an accurate measure of current smoking status in patients with malignant or non-malignant respiratory disease. Smoking behaviour was assessed by self-report and by analysis of whole blood for cotinine, a biomarker of exposure to cigarette smoke, in 166 patients attending a bronchoscopy clinic. Cotinine levels ranged from 2.5 to >400 ng ml−1 blood and were higher in self-reported current smokers (173±123 ng ml−1) than in never smokers (3.7±8.7 ng ml−1) or ex-smokers (20.5±49.0 ng ml−1). Cotinine levels in self-reported current smokers increased with the numbers of cigarettes smoked (p=0.06), and levels in smokers and ex-smokers decreased with the reported length of time since the last cigarette (p=0.001). Using a cotinine level of 20 ng ml−1 and self-report as the gold standard, the sensitivity and specificity for defining current smoking status were 90.2% and 82.4%, respectively. Out of a total of 125 self-reported current non-smokers, 23 (18.4%) had cotinine levels greater than 20 ng ml−1. Smoking prevalence was significantly underestimated by self-report (24.7%) when compared with that defined using blood cotinine levels (36.1%: p<0.001). Misclassification of current smoking status was particularly high in ex-smokers, in patients without malignant respiratory disease, in men, and in those below the median age. Such differential misclassification may result in bias in studies examining associations between current smoking habits and disease risk.

Comparing the Effects of Continuous and Discrete Covariate Mismeasurement, with Emphasis on the Dichotomization of Mismeasured Predictors

It is well known that imprecision in the measurement of predictor variables typically leads to bias in estimated regression coefficients. We compare the bias induced by measurement error in a continuous predictor with that induced by misclassification of a binary predictor in the contexts of linear and logistic regression. To make the comparison fair, we consider misclassification probabilities for a binary predictor that correspond to dichotomizing an imprecise continuous predictor in lieu of its precise counterpart. On this basis, nondifferential binary misclassification is seen to yield more bias than nondifferential continuous measurement error. However, it is known that differential misclassification results if a binary predictor is actually formed by dichotomizing a continuous predictor subject to nondifferential measurement error. When the postulated model linking the response and precise continuous predictor is correct, this differential misclassification is found to yield less bias than continuous measurement error, in contrast with nondifferential misclassification, i.e., dichotomization reduces the bias due to mismeasurement. This finding, however, is sensitive to the form of the underlying relationship between the response and the continuous predictor. In particular, we give a scenario where dichotomization involves a trade-off between model fit and misclassification bias. We also examine how the bias depends on the choice of threshold in the dichotomization process and on the correlation between the imprecise predictor and a second precise predictor.

Possible information bias in a waterborne outbreak investigation

Although an investigation of a cryptosporidiosis outbreak in 1994 in Clark County, Nevada, concluded that illness was associated with consumption of municipal water, no water treatment deficiencies or breakdowns and no water quality changes were discovered during the investigation. We evaluated the strength of the evidence for waterborne transmission and conducted a sensitivity analysis to define the limitations of the epidemiological data. Our analyses suggest a spurious inference of waterborne transmission might be due to differential misclassification bias. If exposure and disease status were incorrectly classified for a relatively small number of study participants, findings of the investigation would be interpreted differently. We offer this example to illustrate the importance of assessing the stability of a relative risk estimate and effect of possible biases during an outbreak investigation.

Is There any Evidence for Differential Misclassification or for Bias Away from the Null in the Swedish Childhood Cancer Study?

Is There any Evidence for Differential Misclassification or for Bias Away from the Null in the Swedish Childhood Cancer Study?

Screening for breast cancer with mammography.

Mammographic screening for breast cancer is controversial, as reflected in greatly varying national policies. To assess the effect of screening for breast cancer with mammography on mortality and morbidity. MEDLINE (16 May 2000), The Cochrane Breast Cancer Group's trial register (24 Jan 2000) and reference lists. Letters, abstracts and unpublished trials. Authors were contacted. Randomised trials comparing mammographic screening with no mammographic screening. Data were extracted by both authors independently. Seven completed and eligible trials involving half a million women were identified. The two best trials provided medium-quality data and, when combined, yield a relative risk for overall mortality of 1.00 (95% CI 0.96-1.05) after 13 years. However, the trials are underpowered for all-cause mortality, and confidence intervals include a possible worthwhile effect as well as a possible detrimental effect. If data from all eligible trials (excluding flawed studies) are considered then the relative risk for overall mortality after 13 years is 1.01 (95% CI 0.99-1.03). The best trials failed to show a significant reduction in breast cancer mortality with a relative risk of 0.97 (95% CI 0.82-1.14). If data from all eligible trials (excluding flawed studies) are considered then the relative risk for breast cancer mortality after 13 years is 0.80 (95% CI 0.71-0.89). However, breast cancer mortality is considered to be an unreliable outcome and biased in favour of screening. Flaws are due to differential exclusion of women with breast cancer from analysis and differential misclassification of cause of death. The currently available reliable evidence does not show a survival benefit of mass screening for breast cancer (and the evidence is inconclusive for breast cancer mortality). Women, clinicians and policy makers should consider these findings carefully when they decide whether or not to attend or support screening programs.

Secondary Leukemia After Adjuvant Chemotherapy for Breast Cancer

Secondary Leukemia After Adjuvant Chemotherapy for Breast Cancer

Cartoon Graphics in the Communication of Accounting Information

The relative merits of schematic faces for the communication of multivariate information have been explored in a number of disciplines. Existing studies in the financial environment suggest that they may be superior to conventional methods in both their communications and decision-making qualities. This paper employs an innovative research design to demonstrate the relative usefulness of schematic faces in the failed/non-failed decision-making context compared to accounting statements and financial ratios. An optimum assignment of financial variables to facial characteristics is suggested, one which demonstrates the usefulness of schematic faces as a decision tool in the financial environment. Schematic faces are shown to be processed more quickly and with no loss of accuracy, compared to more traditional means of presenting financial information. Existing applications of schematic faces have been deficient in a number of areas. This study employs a complex research design involving multiple treatments to overcome the confounding problems of subject variability while also addressing the impact of differential priors and differential misclassification costs.

Carbohydrates and risk of stomach cancer in Uruguay

Carbohydrates and risk of stomach cancer in Uruguay

Comparison of Self-Reported and Physician-Reported Antidepressant Medication Use

PURPOSE: Self-reported medication histories obtained in pharmacoepidemiologic case-control studies are subject to non-differential misclassification and to recall bias. The accuracy of self-reported antidepressant medication use has never been evaluated, but it is important in light of the hypothesis that antidepressant medications may be associated with cancer risk. METHODS: Within a case-control study of several cancer sites, we compared self-reported antidepressant medication use with antidepressant use recorded in physicians’ records. All female cases (n = 147) and controls (n = 119) who reported antidepressant medication use, and a 10% random sample (n = 114) of those who reported no antidepressant use, were asked to provide consent to contact, and the name(s) of their physician(s). These physicians completed a data abstraction form including information on antidepressant prescriptions recorded in patients’ medical records. RESULTS: Substantial agreement was found between subject- and physician-reported antidepressant medication use (kappa = 0.60 [95% confidence interval (CI), 0.47–0.74]; agreement = 80%), and use of specific antidepressant medications (agreement ranged from 82 to 100%), while moderate agreement was observed for duration of use (weighted kappa = 0.56 (95% CI, 0.32–0.79)), and date of first use [weighted kappa = 0.48 (95% CI, 0.23–0.72)]. The level of agreement did not differ markedly between cases and controls, except for duration of use, where agreement was somewhat greater for cases. CONCLUSIONS: The similar level of agreement among cases and controls suggests that differential misclassification (e.g., recall bias) is unlikely in the reporting of most aspects of antidepressant medication use by women. Furthermore, the overall accurate self-reporting of antidepressant use suggests that there should be minimal non-differential antidepressant exposure misclassification.

Effects of misclassification on estimates of relative risk in family history studies

Two types of misclassification that commonly occur in family-genetic studies are distinguished: 1) nondifferential misclassification, in which the probability of error as to phenotype (presence or absence of psychiatric disorder) does not depend on exposure status (being kin to a case or control proband) and 2) differential misclassification, in which it does. Nondifferential misclassification of phenotype reduces the observed relative risk towards the null value, sometimes quite dramatically. Differential misclassification can bias the observed relative risk in either direction, depending on the different values of sensitivity and specificity among relatives of cases and controls. The impact of these biases on genetic-epidemiologic studies is reviewed and discussed. In particular, the ability to detect major gene effects from the pattern of relative risks in first-, second-, and third-degree relatives can be severely compromised. Although there are some methods available to correct the effects of nondifferential misclassification, a major priority for family history studies is to minimize differential misclassification. Genet. Epidemiol. 16:368–381, 1999. © 1999 Wiley-Liss, Inc.

Effects of misclassification on estimates of relative risk in family history studies.

Two types of misclassification that commonly occur in family-genetic studies are distinguished: 1) nondifferential misclassification, in which the probability of error as to phenotype (presence or absence of psychiatric disorder) does not depend on exposure status (being kin to a case or control proband) and 2) differential misclassification, in which it does. Nondifferential misclassification of phenotype reduces the observed relative risk towards the null value, sometimes quite dramatically. Differential misclassification can bias the observed relative risk in either direction, depending on the different values of sensitivity and specificity among relatives of cases and controls. The impact of these biases on genetic-epidemiologic studies is reviewed and discussed. In particular, the ability to detect major gene effects from the pattern of relative risks in first-, second-, and third-degree relatives can be severely compromised. Although there are some methods available to correct the effects of nondifferential misclassification, a major priority for family history studies is to minimize differential misclassification.

Association between obesity and high blood pressure: reporting bias related to gender and age.

To examine the validity of self-reported information on obesity and high blood pressure (HBP) in relation to gender and age, and to explore the impacts of their misclassification on the association between obesity and HBP. Community based cross-sectional study. 1791 adult subjects living in Humboldt, Saskatchewan, Canada. Objectively measured HBP was positive if systolic blood pressure (BP) was > or = 140 mm Hg, diastolic BP was > or = 90 mm Hg or the subject was currently using antihypertensive medication. Self-reported HBP was positive if the subjects gave an affirmative response to the question: 'Has a doctor ever said you had high blood pressure?' Body mass index (BMI) was calculated as weight (kg)/height (m)2. Obesity was defined as a BMI > 27 kg/m2. Measured obesity and reported obesity were based on measured and self-reported information on height and weight, respectively. The sensitivity of self-reported HBP was low, and was lower for men than for women, and for younger subjects than for older subjects. The specificity was similar for both genders. Obese individuals had higher sensitivity and lower specificity than non-obese individuals. The differential misclassification of self-reported HBP caused a bias away from the null when the relative risk for HBP in relation to obesity was estimated. As a result of the gender- and age-related misclassification of self-reported HBP, the modification role of gender and age on the association between obesity and HBP could be altered. The bias caused by self-reported obesity was relatively small and was either toward or away from the null.

Differential misclassification and the assessment of gene-environment interactions in case-control studies.

In case-control studies of interactions between genetic and environmental exposures, differential misclassification of the environmental exposure with respect to disease status can introduce spurious heterogeneity of the stratum-specific odds ratios. In this paper, the authors identify conditions under which differential misclassification does not introduce bias in the interaction parameter when no multiplicative interaction is present, and it biases the interaction parameter toward the null value when a multiplicative interaction is present. The conditions are that (i) conditional on potential confounders, the environmental exposure is independent of the genotype among the controls, and (ii) misclassification of the environmental exposure is nondifferential with respect to the genotype. These conditions can be tested from the misclassified data in the control group, since a test of the independence of the genotype and the misclassified environmental exposure among the controls is a test of the joint hypothesis that conditions (i) and (ii) are both true. Therefore, the authors propose a two-step test for interaction which first tests conditions (i) and (ii) and then goes on to test for interaction, provided the first step hypothesis is not rejected. A summary test procedure to test for gene-environment interactions in the presence of misclassification, based on both a conventional test for interaction and the two-step test, is recommended, and is illustrated with data from a case-control study of the role of diet as a modifier of the association between a metabolic polymorphism and lung cancer.

Differential Exposure Misclassification in Case-Control Studies of Environmental Tobacco Smoke and Lung Cancer

We review some of the literature on the effects of exposure misclassification on the statistical analysis of case-control studies. In particular, we focus on evidence for exposure misclassification which may be different for cases and controls in studies of environmental tobacco smoke (ETS). For example, such misclassification could induce relative risk estimates below unity for low exposure which appears to be the case in U.S. lung cancer and ETS studies. We describe procedures for systematically examining the sensitivity of dose-response statistics on exposure misclassification. These procedures demonstrate how p-values for the null hypothesis of no dose-response trend could be adjusted to account for exposure misclassification. The adjustment procedures were applied to an example based on a recently published large study of ETS and lung cancer in which a p-value for trend was reported as 0.03. In this example it is seen that modest differential exposure misclassification can induce substantial increases in the actual p-value, changing what appears to be statistically significant to decidedly nonsignificant.

National incidence of smoking and misclassification among the U.S. married female population

Because of a lack of representative data on smoking status misclassification among U.S. married females, a two-part study was conducted. Part I was conducted to obtain nationally representative estimates of the percentage of U.S. women who report themselves to be current, former, and never smokers, to determine the concordance of smoking habits among spouse pairs, and to establish field quotas and probability weightings for Part II. Part II was conducted to determine smoker misclassification rates using salivary cotinine as an indication of active smoking. Part I, conducted in January 25–29, 1992, utilized random-digit dialing telephone inter-viewing throughout the 48 contiguous United States. Part II, conducted from February 19, 1992 to March 7, 1992, was a mall-intercept study in nine geographically disperse U.S. cities and it involved interviewing and saliva collection. Among married U.S. women, 25% reported they were current smokers, 22% reported they were former smokers, and 53% reported they were never smokers. Using a cotinine concentration of either >35 ng/ml or >106 ng/ml to indicate regular smoking, 3.61% and 2.55% of regular smokers, respectively, reported themselves to be never smokers. The concordance ratio, an important parameter in correcting for non-differential misclassification bias, was found to be 5.52. In addition, an indication of substantial differential misclassification was found between exposed and unexposed populations. This type of misclassification bias has previously not been accounted for in the adjustment of epidemiology-based risk assessments of tobacco smoke exposure and lung cancer. Taken together, these data suggest that misclassification bias alone is likely to explain any lung cancer risk elevation observed in the U.S. epidemiology of environmental tobacco smoke exposure among non-smoking women.

Medical record validation of maternally reported birth characteristics and pregnancy-related events: a report from the Children's Cancer Group.

Epidemiologic studies frequently obtain exposure information through subjects' self-report (personal interview or mailed questionnaire). The authors used data from a case-control study of infant leukemia, to assess the validity and reliability of maternally reported information on birth characteristics such as birth weight, reproductive history, and medical procedures. Cases were gathered from the Children's Cancer Group, a United States and Canadian cooperative clinical trails group with approximately 100 member and affiliate institutions, during 1983-1988. Telephone interviews were completed for 302 cases and 558 matched controls. Medical records of the index pregnancy were obtained for 287 cases and 467 controls. Correlations between medical charts and maternal interview were high for birth weight (r = 0.98, kappa = 0.9) and gestational age (r = 0.86, kappa = 0.6). Mean differences between the two sources were small, -10.5 g for birth weight and -0.36 weeks for gestational age. Reproductive history and medical procedures had high to moderate reliability. Problems after delivery and pregnancy complications generally had low validity and reliability. Little evidence of differential misclassification was found. Time between delivery and interview ranged from zero to 8 years and did not greatly affect reliability. This study suggests that validity and reliability of maternally reported pregnancy and delivery information may differ with the nature of the factor of interest, but is affected little by time from birth or case-control status.

Monte Carlo modeling of epidemiological studies

Abstract As epidemiologists search for smaller and smaller effects, the statistical uncertainty in their studies can be dwarfed by biases and systematic uncertainty. We here suggest that Monte Carlo techniques are very useful to estimate some of these biases and uncertainties, and perhaps to avoid them entirely. We illustrate this by two simple Monte Carlo simulations. First, we show how often false positive findings, and sometimes false negative findings, can result from 33 differential misclassification of the exposure status. Secondly, we show how a bias, that we call “the binning bias,” can be caused if the investigator chooses bin boundaries after he has seen the data. We show how an allowance might be made for such a bias by increasing the uncertainty bounds. This would put the presentation of the results on a par with the presentation in physical sciences where a quantitative estimate of systematic errors is routinely included with the final result. Finally, we suggest how similar Monte Carlo simulations carried out before and during the study can be used to avoid the biases entirely.