Examine Subsets of the Population with Differing Exposure Data Quality

Holistic Hormone Balance

Natural Menopause Relief Secrets

Get Instant Access

Accuracy of exposure assignment may vary in predictable ways in relation to attributes such as gender, education, or age. Based on previously reported research or based on validation studies contained within the study of interest, groups may be defined for whom the routine exposure measure is more likely to be valid than it is for other groups in the study. Younger persons may show a greater or lesser concordance of the routine and superior exposure measures than older persons, or women may report more or less accurately than men. Often, groups expected to have superior cognitive function, e.g., younger participants versus the elderly, or those without a history of alcohol and drug abuse compared to those with such a history, are likely to provide better quality information through recall and self-report. In fact, when the disease itself is associated with cognitive decline, e.g., studies of chronic neurodegenerative disease, case-control studies are susceptible to bias because cases provide data of inferior accuracy relative to controls. Some individuals in the study may have experienced the etiologically relevant period in the more distant past than others. All other considerations equal, those who are reporting for a more recent calendar period may well provide better data than those reporting for a more remote time.

Physiologic differences among individuals may also make the exposure measure more accurate for some than others. Differences arising from genetic variation or induced metabolic changes can create variation in the biologically effective dose for a given level of exogenous exposure. Depending on the particular exposure of interest, a variety of hypotheses might be put forward regarding subgroups of participants in whom the accuracy of data would be higher or lower. For example, in addressing the question of whether historical exposure to organochlorines such as DDT and PCBs may be related to the development of breast cancer, a major challenge is in accurately reconstructing historical exposure. In case-control studies, in particular, measurements of present-day serum residues of the chemicals of interest serve as an exposure indicator for lifetime exposure history. While there are a number of factors that influence the changes in body stores and serum levels over time, lactation is a primary means by which such organochlorines are excreted. All other considerations equal, serum organochlorines in women who have lactated are less indicative of long-term historical exposure levels than for women who have not.

This understanding of lactation as a modifier of exposure was exploited in recent studies of organochlorines and breast cancer to define strata in which the present-day serum markers are more valid markers of long-term exposure (Mil-likan et al., 2000). The association between serum levels of DDE and PCBs were examined in relation to breast cancer in strata of women who were nulliparous, parous but never breastfed, and those who were parous and had breastfed (Table 8.3). These data suggest a weak positive association, though without a dose-response gradient, limited to women who had never breastfed, i.e., the first two strata. Among those who had breastfed, the odds ratios were close to or some-

Table 8.3. Odds Ratios for Lipid-Adjusted DDE and PCBs and Breast Cancer, Stratified by Parity and History of Breastfeeding, North Carolina, 1993-1996





Parous, Never Breastfed

> 1.044 134 111 Total PCBs

Parous, Ever Breastfed

> 1.044 74 92 Total PCBs

*Adjusted for age, age-squared, and race.

fAdjusted for age, age-squared, race, menopausal status, BMI, body mass index, HRT, hormone replacement therapy use, and income. tpp'-DDE in pg/g lipid. §Total PCBs in pg/g lipid. Millikan et al., 2000.

what below the null. Such results, though suggestive at most, may be reflective of the superiority of the serum measures as an exposure indicator for women who had never breastfed and accurately reflect a small increase in risk associated with the exposure.

As illustrated by the above example, information on predictors of accuracy in exposure classification can be used to create homogeneous strata across which the validity of exposure data should vary in predictable ways. All other influences being equal, those strata in which the exposure data are better would be expected to yield more accurate measures of association with disease than those strata in which the exposure data are more prone to error. Identifying gradients in the estimated validity of the exposure measure and examining patterns of association across those gradients serves two purposes—it can provide useful information to evaluate the impact of exposure misclassification and also generate estimates for subsets of persons in whom the error is least severe. Note that it is not helpful to adjust for indicators of data quality as though they were confounding factors, but rather to stratify and determine whether measures of association differ across levels of hypothesized exposure data quality.

The quality of women's self-reported information on reproductive history and childhood social class was evaluated in a case-control study of Hodgkin's disease in northern California using the traditional approach of reinterviewing some time after the initial interview (Lin et al., 2002). Twenty-two cases and 24 controls were reinterviewed approximately eight months after their initial interview, and agreement was characterized by kappa coefficients for categorical variables and intraclass correlation coefficients for continuous measures. Whereas cases and controls showed similar agreement, education was rather strongly associated with the magnitude of agreement (Table 8.4). Across virtually all the measures, women who had more than a high school education showed better agreement than women of lower educational level, suggesting that the more informative results from the main study would be found within the upper educational stratum.

Non-specific markers of exposure data quality such as age or education may also yield strata that differ in the magnitude of association for reasons other than the one of interest. There may be true effect measure modification by those attributes, in which exposure really has a different impact on the young compared to the old, or there may be other biases related to non-response or disease mis-classification that cause the association to differ. When the identification of persons who differ in the quality of their exposure assignment is based on specific factors related to exposure, such as having been chosen for a more thorough protocol or having attended a clinic or worked in a factory in which more extensive exposure data were available, then the observed pattern of association is more likely to be reflective of the accuracy of the exposure marker as opposed to other correlates of the stratification factor.

Table 8.4. Kappa or Intraclass Correlation Coefficients Among Subjects (n = 46) Reinterviewed between 1992 and 1995 in a Case-Control Study of Hodgkin's Disease, Stratified by Education, Northern California

Kappa or Intraclass Correlation Coefficient*



Age at first period

0.827 (n =


0.920 (n =


Age at first period^

0.541 (n =


0.848 (n =


Number of pregnancies

0.584 (n =


0.823 (n =


Number of live births

0.632 (n =


0.887 (n =


Use of birth control pills or shots

0.769 (n =


0.877 (n =


Number of playmates at age 8

0.158 (n =


0.779 (n =


Birth weight in pounds

0.943 (n =


0.966 (n =


History of mononucleosis

0.000 (n =


0.907 (n =


Mean reliability^



(95% CI)

(0.345, 0.735)

(0.837, 0.912)

♦Calculation of kappa and intraclass correlation coefficients does not include missing or unknown responses.

¿Calculated by subtracting year of birth from reported year at first period.

¿Bootstrapped difference between means (more than high school-high school or less) and 95% CI: 0.319 (0.147, 0.521). CI, confidence interval. Lin et al., 2002.

♦Calculation of kappa and intraclass correlation coefficients does not include missing or unknown responses.

¿Calculated by subtracting year of birth from reported year at first period.

¿Bootstrapped difference between means (more than high school-high school or less) and 95% CI: 0.319 (0.147, 0.521). CI, confidence interval. Lin et al., 2002.

Was this article helpful?

0 0
10 Ways To Fight Off Cancer

10 Ways To Fight Off Cancer

Learning About 10 Ways Fight Off Cancer Can Have Amazing Benefits For Your Life The Best Tips On How To Keep This Killer At Bay Discovering that you or a loved one has cancer can be utterly terrifying. All the same, once you comprehend the causes of cancer and learn how to reverse those causes, you or your loved one may have more than a fighting chance of beating out cancer.

Get My Free Ebook

Post a comment