Martha M. Werler
One challenge in studying environmental exposures related to oral cleft risks is assessment of exposure. This chapter provides a general discussion of study design issues and how they relate to the assessment of exposures. There is a brief description of the consequences of inaccurate assessment and approaches to measuring it. Then, several sources of exposure information are described. Following is a section describing the methods of assessing specific exposures that are relevant to the study of oral clefts.
Study designs
Accurate assessment of exposures depends, to a large extent, on the study design. As described in Chapter 9, the cohort and case-control observational designs are the most widely used, in addition to the cross-sectional method. Cohort studies typically measure exposure status prospectively before the outcome is known. Prospective studies are considered optimal for assessment of exposures. In the case of oral clefts, prospective studies involve identifying a cohort of pregnant women, measuring exposure status, and following the women to the end of pregnancy to determine the presence or absence of oral clefts in their offspring. Case-control studies identify cases with the outcome of interest and a control group, then measure exposure status, typically, retrospectively for the etiologically relevant time period. Accurate retrospective assessment is a major challenge in case-control studies, but certain steps can be taken to maximize accuracy. Crosssectional studies identify cases and controls in similar fashion to that in a retrospective study, but exposure status is measured at the time of case/control identification or later. Unless maternal exposure status after delivery reflects the status early in pregnancy, inferences made from cross-sectional findings are severely limited.
Although oral clefts are one of the most common birth defects, their occurrence at birth (1.7/1000) is rare enough to complicate their study. Since it is most appropriate to separate cleft lip with or without cleft palate from cleft palate alone (see Chapter 1,2, and 5), prevalence rates are further reduced to 1.1/1000 births and 0.6/1000 births, respectively. Hence, prospective studies must follow approximately 50,000 pregnant women to identify 85 infants with cleft lip with or without cleft palate and 30 infants with cleft palate alone. The greater efficiency of the case-control design explains why the majority of oral cleft studies are retrospective. However, the gain in efficiency of retrospective studies is countered by difficulties in accurate exposure assessment.
Exposure assessment
Misclassification
The main goal of conducting studies environmental exposures in relation to oral clefts is to estimate risk. As exposure is more accurately assessed, so the risk measure is more accurately estimated. When exposure status is poorly ascertained across all study subjects and large portions of subjects are misclassified, the net result is usually underestimation of risk. In other words, in the presence of random misclassification, risk measures (e.g., odds ratios or relative risks) are biased toward the null value. When exposure status is poorly ascertained for one of the study groups compared to the other, known as differential misclassification, the effect on risk measures may be great but the direction (under- or overestimation) is more difficult to predict. Both random and differential misclassification can be present in any of the three previously described study designs. However, there is particular concern over differential misclassification in retrospective studies based on self-reported exposure information, referred to as recall bias.
Recall Bias
Concern about recall bias stems from the intuitive idea that a mother who gives birth to a child with an oral cleft may reflect back on her pregnancy, searching for a reason to explain why her baby is affected. Indeed, many a clinician can relay stories of concerned mothers asking whether one exposure or another in pregnancy could be ‘the reason’. Under this scenario, case mothers would be more accurate reporters than mothers of ‘normal’ children. Conversely, the reverse situation can be invoked, where mothers of cases might deny exposures that women are warned to avoid in pregnancy (e.g., alcohol) and reporting accuracy would be greater for controls than cases. There is little empirical evidence of recall bias in studies of oral clefts, not because there is evidence that it does not exist but, rather, because it is difficult to measure. To measure recall bias, a gold standard must be available against which exposure status is compared. For example, retrospective maternal reports might be compared to some other source of exposure information such as stored biologic samples or documentation in medical records. However, such sources are often not available. Further, medical records may provide an inaccurate assessment of exposure and, thus, a poor gold standard (Drews et al., 1990; Werler et al., 1989). Prospective studies in which information was collected from women both during and after pregnancy (MacKenzie and Lippman, 1989; Klemetti and Saxen, 1967) have been important in measuring accuracy of recall, but they offer less in terms of assessing recall bias. This is due to the rarity of oral clefts and other malformations. The gold standard (whether prospective interview data or some other source) would need to be available on hundreds of thousands of pregnant women to compare postpartum recall accuracy between mothers of oral cleft cases and mothers of normal infants for exposures that are rare (<10%). Further, it is likely that the accuracy of recall and the potential for recall bias would vary across exposures, depending on the respondent's attitude, society's view, or whether the media had recently cast any negative attention on it. Also, increased length of time between interview and early pregnancy is thought to negatively affect recall. Due to concerns of recall bias, some retrospective studies have used a control group comprised of infants with other major structural malformations, assuming that accuracy of recall would be similar between mothers of oral cleft cases and mothers of infants with other major malformations. Such a study design requires that the exposure(s) under study not be associated with the malformations in the control group (Swan et al., 1992), to avoid introducing a selection bias. Since retrospective studies are often launched with the intent of gathering information on a wide variety of exposures, it may be difficult to predict which exposures are vulnerable to recall bias and which are associated with non-oral cleft malformations. If resources allow, both non-malformed and malformed control groups may be utilized, to address concerns of potential recall and selection biases. Studies have shown that risk estimates for various exposures do not vary when malformed and non-malformed controls are used (Werler et al., 1996,1999; Khoury 1994). In any event, the best approach is to obtain the most accurate exposure information possible for all study subjects, to reduce the likelihood of either random or differential misclassification.
Validity Assessment
Despite the difficulties described above in determining the presence of recall bias or differential misclassification, there is considerable merit in determining the validity of measuring exposures in both prospective and retrospective studies (Drews and Greenland, 1990). Validity assessment involves comparing the exposure measure of interest to a gold standard. One measure validity is sensitivity, or the probability of being classified as exposed if truly exposed. Conversely, specificity is the probability of being classified as not exposed if truly unexposed. Using the data presented in Table 10.1 as an example, smoking measured by interview is compared to cotinine levels in first-trimester urine samples (as the gold standard). Sensitivity would be the proportion of cotinine identified-smokers who were classified by interview as smokers (25/30 = 83%); specificity would be the proportion of women identified by cotinine levels as nonsmokers who were also classified by interview as nonsmokers (69/70 = 99%).
|
TABLE 10.1. Validity Assessment of Cigarette Smoking |
|||||||||||||||||||
|
|||||||||||||||||||
Exposure information may be available from two separate sources or measures, but one may not necessarily be considered a gold standard. In this instance, the amount of agreement or correlation between the two exposure measures can be assessed. Simple percent agreement can be misleading because a certain amount of agreement occurs by chance. For dichotomous factors, the Kappa statistic corrects for the amount of agreement that is expected to occur by chance (Thompson and Walter, 1988; Fleiss, 1973). The Kappa statistic is considered to represent excellent agreement when greater than 0.75, fair to good agreement in the range of 0.4 to 0.75, and poor agreement when less than 0.4. However, it is particularly dependent on the prevalence of exposure, tending to approach low values when prevalence rates are low, regardless of sensitivity and specificity.
It has been recommended that sensitivity analyses be employed to assess the likelihood that an observed association between exposure and disease might be due to measurement error. This is a theoretical exercise where effect estimates (e.g., odds ratios) are calculated after assuming certain amounts of error in exposure measurement (i.e., a particular sensitivity value). For example, a study of cigarette smoking in pregnancy, based on birth certificate data, observed 35% of cases exposed and 20% of controls exposed and produced an odds ratio of 2.2 (Table 10.2).
If the sensitivity of the observed cigarette smoking data was assumed to be higher for cases than controls (80% vs. 60%, as might occur in the presence of recall bias), the odds ratio would decrease to 1.6. If the sensitivity disparity were greater, say 70% vs. 40%, the odds ratio would decrease to 1.0. However, in the former scenario, the prevalence of smoking among controls would be 50%, which is unreasonably high based on reported smoking and cotinine measures in pregnant women (Kendrick and Merritt, 1996; Haddow et al., 1987). Hence, it would be reasonable to assume that extreme differential recording of cigarette smoking would not completely account for the increased risk in the example.
|
TABLE 10.2. Sensitivity Analysis |
|||||||||||||||||||||||||||||||||||||||||||||
|
|||||||||||||||||||||||||||||||||||||||||||||
Gestational Timing
As described in detail Chapter 1, fusion of the upper lip and palate occurs within the first 10 weeks of gestation. Thus, epidemiologic studies should aim at capturing exposure status early in pregnancy. To do so, it is necessary to establish gestational timing of pregnancy. The usual approach is based on the last menstrual period. However, approximately 5% of women have unreliable dates due to irregular cycles or missed menstrual periods. Sonograms in the first half of pregnancy accurately assess gestational timing, with a general rule being that the earlier in gestation sonogram is performed, the more accurate the dating (Mongelli et al., 1996; Hall, 1990). Other methods of gestational dating, such as fundal height of the mother and Dubowitz exam of the infant, are considered less reliable (Hall, 1990).
Sources of Information
Biologic Specimens
There is a range of possible biologic specimens that can serve as potential markers of exposure, including amniotic fluid; placenta; maternal serum, urine, hair, and nails; and infant serum and urine. Each tissue has its own set of advantages and disadvantages. For example, maternal hair or nail samples have the advantage of representing cumulative exposure levels in previous months, thus allowing the time of sampling to follow the etiologically relevant time period; however, only limited exposures can reliably be measured from hair, such as trace elements and cocaine, and contamination is not uncommon (Hunter, 1990). Maternal cumulative exposure may not be the best indicator of fetal exposure. For example, maternal blood measurements of compounds that are stored in fat or bone represent circulating levels and may more closely represent fetal exposure than measurements of storage tissues, which would represent maternal cumulative exposure. However, blood samples would need to be collected during early pregnancy, which is not always feasible. When the exposure of interest is a one-time event, such as acute exposure to a toxin resulting from an environmental disaster, then compounds with long half-lives can be measured in biologic tissues years later, allowing identification of exposed and unexposed subjects retrospectively. Exposures that are likely to change during pregnancy (e.g., diet, cigarette smoking, coffee consumption, and alcoholic beverage drinking) can be more difficult to accurately measure in biologic samples from a feasibility standpoint in that the timing of measurement should coincide with the etiologically relevant time period, i.e., early gestation.
Biologic measures may reflect intake amount or levels in tissues but not always both. In other words, intake amounts may not correlate with tissue levels because other factors (e.g., genetic factors, bioavailability, other exposures) may play a role. Both intake amounts and tissue levels are of interest; the former is potentially modifiable, and the latter is more closely related to risk (Hunter, 1990).
Environmental or Occupational Sampling
Testing of exposures in the environment can be done to measure air pollutants, water contaminants, or soil contaminants. These measures may not represent individual exposure levels; thus, they are indirect or ecologic measures. Occupational records are sometimes available as either indirect or direct measures of exposure. An example of an indirect measure of occupational exposure would be employment in an industry or building known to use solvents, thereby indicating the potential for exposure. Direct measurement would indicate exposure status per individual rather than per residence, employment category, or site. Because direct measurement of environmental exposures can be difficult, requiring personal monitoring devices, indirect measures are more often used. However, the accuracy of indirect measures can be improved by collecting information on individual behaviors that may influence the likelihood of exposure. For example, the number of days of employment in a particular occupation as an indirect measure of an occupational exposure could be enhanced by asking employees if they wore protective devices. Another example is that individual residences linked to water distribution records (as an indirect measure of exposure to contaminated water) could be enhanced by asking individuals about consumption of tap vs. bottled water. Environmental databases on air pollution dispersion, pesticide use patterns, water pollution, and toxic waste sites can be accessed for geographic information systems studies (Rodenbeck et al., 2000; Moore and Carpenter, 1999). Also, more involved dosimetric models can be implemented which incorporate environmental, individual, and biologic data (Hatch and Thomas, 1993). Use of these sophisticated models can help to characterize potential exposure, but without information on individual exposure factors, they are likely to fall short of accurate assessment.
Records
Medical records can be an excellent source for exposures or procedures that are accurately and routinely documented in a standardized manner. For example, infertility procedures or anesthetic exposures are usually well documented in medical records. Other exposures, such as vitamin use, over-the-counter medication use, or infectious diseases, may not be well documented. For prescription medications, pharmacy records are a reasonable source, assuming the woman was compliant in her use.
Vital records are sometimes used as a source of information. Birth certificates include demographic information, such as maternal age, race, and birth order. In most U.S. states, information is also available on maternal exposures such as cigarette smoking, alcohol consumption, and illicit drug use during pregnancy. However, this source is considered inaccurate for such exposures, and details on level timing of exposure are typically not recorded (Dietz et al., 1998).
Maternal Report
Finally, maternal report is a common source of exposure information. Reported exposures rely on maternal recall and, therefore, are subject to error. Studies have shown that such errors and the resulting random and differential misclassification can be reduced by asking standardized and detailed questions (Mitchell et al., 1986; Klungel et al., 2000; Eskenazi and Pearson, 1988). Questions that are open-ended or that lead the respondent should be avoided. In addition, if questions on a particular exposure or set of exposures can be asked in several ways, it has been shown that reporting increases. For example, information on medication use can be elicited by indication (“Did you take anything for headache?”), type of drugs (“Did you take any pain- or fever-reducing medications?”), and specific drugs (“Did you take any aspirin, acetaminophen, or ibuprofen?”). In addition, questions can be tailored to elicit details on the most relevant information, i.e., exposure timing (onset and duration), dose, frequency, changes in use, gestational timing, and potential confounders.
The method of administering a questionnaire is related to the quality of information. Self-administered questionnaires depend on the literacy and sometimes intellect of the respondent. Further, questions may be unintentionally skipped or left incomplete (O'Toole et al., 1986). Both telephone and in-person interviews can overcome these limitations, with higher-quality information obtained in person (Patterson et al., 1998; Hermann, 1985; Weeks et al., 1983; Einarson 1999). In addition, participation rates tend to be higher when study subjects are verbally asked to participate (Weeks et al., 1983).
Specific Exposures
Cigarette Smoking
There are several biomarkers of exposure to cigarette smoke; but cotinine, the main nicotine metabolite, is at present the most stable, sensitive, and specific marker for either individual exposure or passive exposure (Benowitz, 1999; Jauniaux et al., 1999). Cotinine can be measured in blood, saliva, urine, and amniotic fluid. Findings on cigarette smoking in relation to oral cleft risk appear to be dose-dependent (number of cigarettes smoked per day); hence, details on quantity of smoking should be collected. However, quantity and timing of smoking in pregnancy are not well documented on birth certificates in the United States, with approximately 30% underreporting compared to urinary cotinine measures (Kendrick et al., 1995). Documentation in prenatal medical records is generally better than on birth certificates; but underreporting remains, and information on quantity and timing is often missing. The quality of questionnaire data on smoking depends on the way questions are asked. Kharrazi et al. (1999) compared four questions on smoking in pregnancy to cotinine levels determined from second-trimester blood samples. Timing-specific questions on cigarette smoking by telephone interview were estimated to be 87% sensitive, whereas a simple yes/no question during a prenatal care visit was only 47% sensitive. Data suggest that pregnant smokers under-report that they smoke at all and that, among those who report smoking, the amount smoked is underreported (Ford et al., 1997).
Alcohol
Maternal alcohol consumption is difficult to accurately measure in early pregnancy. Various biologic markers of alcohol consumption have been examined, including mean red cell volume, γ-glutamyl transpeptidase, acetaldehyde in whole blood, and carbohydrate-deficient transferrin, each with different sensitivities and specificities. Stoler et al. (1998) used all four markers as a screen for alcohol abuse in pregnant women and found greater sensitivity than with each marker alone, using abnormal phenotype in offspring as the gold standard. Several different screens, such as TWEAK, T-ACE, Michigan Alcohol Screening Test or MAST, and CAGE (Russell et al., 1996), have been developed to identify heavy alcohol consumers with reasonably high sensitivity and specificity (>75%). However, moderate and occasional binge alcohol drinking are more common behaviors and may carry risks for oral clefts (Werler et al., 1991; Munger et al., 1996). To measure these more casual or intermittent patterns of use, separate questions on average and binge frequency (how often), timing (when), dose (drinks per day), and type (beer, wine, liquor) should be asked in detail. One pitfall in collecting alcohol information is asking separately about the frequency and dose of beer, wine, and liquor intake. Data collected in this manner cannot accurately assess the total frequency and dose without assuming that only one type of beverage is consumed on any given day. For example, a woman may report drinking two glasses of wine per drinking day 3 days per week and two liquor drinks per drinking day 3 days week. If the woman drank only 3 days per week, her daily dose would be four drinks; if she drank 6 days per week, her average daily dose would be two drinks. Sensitivity and specificity values have not been estimated for various questions on patterns of use, but presumably sensitivity is not high and specificity is not low.
Medications and Illnesses
Various medications and illnesses, such as antiepileptics (Abrishamchian et al., 1994) and the common cold (Zhang and Cai, 1993), have been suggested to affect oral cleft risk; and other such exposures deserve to be studied. Information on prescription medications can be obtained from medical records, but one must assume that the woman followed her clinician's orders in a timely and otherwise compliant fashion. Pharmacy records are one step closer in that they indicate that the prescription was filled, but one must assume that the woman took the medication as prescribed. Also, a woman may have taken a medication that was not documented in her medical record or was filled at a different pharmacy. Medical records and pharmacy sources are most useful for cohorts of pregnant health maintenance organization members, where different medical and pharmacy services are documented within one system. These sources have the added benefit of being documented during pregnancy and are not subject to recall bias. However, most medications taken in pregnancy are not prescribed (Werler and Mitchell, 2000) and are more difficult to ascertain because they tend not to be documented in medical records. The most common approach to assessing exposure to over-the-counter medications is by interview. Questions must be detailed on specific product, frequency, dose, timing, and indication for use. To assess confounding by underlying illness in drug analyses, it is necessary to obtain information on the illness(es) among unmedicated women as well. Also, depending on the drug, route of administration may be important; e.g., corticosteroid exposure includes oral, inhaled, and topical routes. For over-the-counter drugs, particularly cough/cold/analgesic products, it is important to identify the product name, type (capsule, packet, liquid), and packaging because product lines are often labeled according to the symptoms they treat and vary in the components they contain. For example, Tylenol® contains acetaminophen and Tylenol Flu® contains acetaminophen, dextromethorphan, and pseudoephedrine. Over-the-counter medications are usually taken intermittently and on a more casual basis, which complicates accurate assessment, presumably reducing both sensitivity and specificity.
Issues surrounding the assessment of illnesses in pregnancy parallel those of medication use in pregnancy. Chronic or severe illnesses are more likely to be reliably documented in medical records than those that are acute or less severe. Infectious illnesses can be extremely difficult to accurately assess in observational studies because it is prohibitive to conduct sequential serology tests in early pregnancy to identify true exposure to any given specific agent. Thus, studies usually rely on reported illness events, such as colds or flu. Assessment of fever is subject to many of the same limitations. In interview studies on illnesses, the quality of information can be improved by asking details on timing, whether the illness was diagnosed by a clinician, what symptoms were experienced, whether medication was taken and what type, and whether fever accompanied the illness (and, if so, fever duration and highest temperature).
Vitamins and Diet
For biologic measurement of various nutrients, serum and plasma are commonly used. However, red blood cells may provide a better indication of nutrient status due to greater stability compared to serum and plasma levels. One potentially useful source of second-trimester blood samples, for which measurements might correlate with early gestation levels, is banked maternal serum following prenatal screening. Long-term storage may result in degradation of nutrients, but such degradation should not be differential for cases and controls. Detailed discussion of sampling and measurement issues can be found elsewhere (Hunter, 1990). Hair and nail samples provide stable measures of heavy metals but are subject to contamination from cleansing and polishing products.
Many studies rely on reported supplement and dietary intakes to assess nutrient status. For supplements, questions must be asked about specific products because there are hundreds of multivitamin products with varying amounts of different vitamin and mineral components. Most women begin use of prenatal multivitamin preparations after pregnancy is clinically recognized (Werler et al., 1999), regardless of whether or not they were routine vitamin takers before pregnancy. Therefore, it is important to ascertain the timing of use of each supplement and to determine exposure status during the developmentally relevant time period.
Approaches to assessing dietary intake include 24 h dietary recalls, food diaries, and food-frequency questionnaires (Witschi, 1990). The first two methods are effective for measuring dietary intake but reflect intake at the time of data collection. Since dietary patterns often change during early pregnancy due to nausea and then change again once nausea has passed, intake at the time of measurement may not reflect the developmentally relevant time period. One advantage of food-frequency questionnaires is that they can refer to a specific time, e.g., pre-pregnancy or early pregnancy. Indeed, food-frequency questions can identify changes in diet during pregnancy (Brown et al., 1996) and have been validated among pregnant women (Wei et al., 1999). However, if there is a long delay between early pregnancy and when dietary intake is measured, recall of dietary patterns may be difficult for the relatively short period of organogenesis (approximately 8 weeks). If there is an interval of several months or more between early gestation and the time that dietary information is collected, food-frequency questions targeted to average diet during the months preceding pregnancy may be better recalled and may better reflect nutrient status of very early gestation.
The validity of food-frequency data collected retrospectively has been examined in several non-pregnant populations (Sobell et al., 1989; Block 1990; Mares-Perlman et al., 1993) but to date has not been reported in postpartum populations as a reflection of early or prepregnancy diet. Among nonpregnant women, correlation coefficients for retrospective measurement of diet compared to 24 h recalls vary across the range of foods from low to high but, on average, are in the moderate range (0.52) (Salvini et al., 1989); for specific nutrients, correlation coefficients ranged from 0.36 for vitamin A without supplementation to 0.75 for vitamin C with supplements (Willett et al., 1985). A detailed discussion of assessment of diet can be found elsewhere (Willett, 1990).
Solvents and Pesticides
Because exposure to solvents and pesticides can come from more than one source, it is important to be as comprehensive as possible in assessments. For example, occupational studies should include assessment of exposure away from the workplace. This requires contact with study individuals, which in some occupational or environmental studies may not be feasible; but if at all possible, the gain in accuracy of exposure assessment may be worth the extra expense. Further, in occupational, environmental, and home settings, exposure to a single product may involve several different chemical agents. Every attempt should be made to determine the exact chemicals of exposure. In occupational settings, records of chemicals used can be assessed. There is an Environmental Protection Agency number (EPA Reg. No.) on all pesticide products sold in the United States. This unique identifier can be easily ascertained from products that are still on hand and then linked to its chemical contents. The typical research goal is to identify whether risks are confined to a particular chemical or group of chemicals, but often there is so much overlap in exposure to individual chemicals that it prohibits estimating independent effects. Nevertheless, there is value in identifying risks associated with mixes of chemicals because public health interventions may occur for an overall exposure scenario rather than for a specific chemical (Hertz-Picciotto, 1998). A detailed discussion of pesticide and solvent exposure measurement can be found elsewhere (Armstrong et al., 1992).
Confounding Factors
While accurate assessment of exposure is important in epidemiologic studies, potential confounding factors also deserve careful attention. In studies of oral clefts, factors that are known to be associated with their occurrence are obvious candidates as confounders. For oral clefts, information should be collected on infant's race; ethnicity; sex; maternal and paternal ages; maternal epilepsy; maternal use of anticonvulsants, corticosteroids, high-dose vitamin A, or trimethoprim; maternal occupation; maternal smoking; and maternal use of alcohol. However, other factors which have not previously been identified may be related to both oral clefts and the exposure under study. It is important to collect accurate information on potential confounders, to reduce the likelihood of spurious estimation adjusted risks (Greenland, 1980).
Likewise, factors that may act as effect modifiers should be accurately measured. Positive family history of oral clefts is one such potential effect modifier. Accurate assessment is best accomplished by contacting family members individually for medical information. One study has documented that family history of birth defects was more accurately obtained by this method as opposed to maternal report (Romitti et al., 1997). Inaccurate measurement can result in an apparent lack of effect modification when in fact it exists or in an apparent interaction when none exists (Greenland, 1980).
Data Collection
It is important to record data in a way that allows clear understanding of exposure status. Because most sources of exposure information are not perfect, allowance must be made for incomplete or missing information. For example, no notation on alcohol consumption in a medical record may represent truly missing information (the question was never asked of the patient) or a negative response (only positive responses were recorded). Likewise, medical records or maternal report may indicate use of cough medication, but the specific agent is missing. In addition, recall of details of illnesses and medication use that occurred several months earlier can be difficult. Data collection instruments should include options such as “missing,” “not otherwise specified,” and “don't know” to accommodate these situations. In maternal interviews, there is a natural tendency to try to force a study participant into an exposed or unexposed category; however, random misclassification will be reduced if truly unclear information is recorded and analyzed as such.
Conclusion
Accurate assessment of exposures is essential for valid study of risk factors for oral clefts. Prospective studies may reduce exposure misclassification by collecting information closer to the etiologically relevant time period (the first trimester), but they are less efficient than retrospective studies. Therefore, the majority of studies of environmental risk factors for oral clefts are retrospective in design and thereby present further challenges in accurate exposure assessment. It is important that efforts be made to collect detailed exposure information in as complete and accurate a fashion as possible, to reduce the likelihood of misclassification. While such efforts can be laborious during the data collection phase of studies, there is a gain in the validity of risk estimation that justifies the added effort.
Acknowledgements
I thank Allen Mitchell for his helpful comments.
References
Abrishamchian, AR, Khoury, MJ, Calle, EE (1994). The contribution of maternal epilepsy and its treatment to the etiology oral clefts: a population based case-control study. Genet Epidemiol 11: 343–351.
Armstrong, BK, White, E, Saracci, R (1992). Principles of Exposure Measurement in Epidemiology. Monographs in Epidemiology and Biostatistics, Vol. 21. Oxford: Oxford University Press.
Benowitz, NL (1999). Biomarkers of environmental tobacco smoke exposure. Environ Health Perspect 107(Suppl 2): 349–355.
Block, G, Woods, M, Potosky, A, Clifford, C (1990). Validation of a self-administered diet history questionnaire using multiple diet records. J Clin Epidemiol 43: 1327–1335.
Brown, JE, Buzzard, IM, Jacobs, DR, Jr, et al. (1996). A food frequency questionnaire can detect pregnancy-related changes in diet. J Am Diet Assoc 96: 262–266.
Dietz, PM, Adams, MM, Kendrick, JS, et al. (1998). Completeness of ascertainment of prenatal smoking using birth certificates and confidential questionnaires: variations by maternal attributes and infant birth weight. Am J Epidemiol 148: 1048–1054.
Drews, CD, Greenland, S (1990). The impact of differential recall on the results of case-control studies. Int J Epidemiol 19: 1107–1112.
Drews, CD, Kraus, JF, Greenland, S (1990). Recall bias in a case-control study of sudden infant death syndrome. Int J Epidemiol 19: 405–411.
Einarson, A, Ahmed, SF, Gallo, M, et al. (1999). Reproducibility of medical information obtained via the telephone vs personal interview. Vet Hum Toxicol 41: 397–400.
Eskenazi, B, Pearson, K (1988). Validation of a self-administered questionnaire for assessing occupational and environmental exposures of pregnant women. Am J Epidemiol 128: 1117–1129.
Fleiss, JL (1973). Statistical Methods for Rates and Proportions. New York: John Wiley & Sons.
Ford, RP, Tappin, DM, Schluter, PJ, Wild, CJ (1997). Smoking during pregnancy: how reliable are maternal self reports in New Zealand? J Epidemiol Community Health 51: 246–251.
Greenland, S (1980). The effect of misclassification in the presence of covariates. Am J Epidemiol 112: 564–569.
Haddow, JE, Knight, GJ, Palomaki, GE, et al. (1987). Cigarette consumption and serum cotinine in relation to birthweight. Br J Obstet Gynaecol 94: 678–681.
Hall, MH (1990). Definitions used in relation to gestational age. Paediatr Perinat Epidemiol 4: 123–128.
Hatch, M, Thomas, D (1993). Measurement issues in environmental epidemiology. Environ Health Perspect 101(Suppl 4): 49–57.
Herrmann, N (1985). Retrospective information from questionnaires. II. Intrarater reliability and comparison of questionnaire types. Am J Epidemiol 121: 948–953.
Hertz-Picciotto, I (1998). Environmental epidemiology. Domain of environmental epidemiology. In: Modern epidemiology, 2nd ed., edited by KJ Rothman and S Greenland. Philadelphia: Lippincott-Raven, pp. 555–583.
Hunter, D (1990). Biochemical indicators of dietary intake. In: Nutritional epidemiology. Monographs in epidemiology and biostatistics, Vol. 15, edited by W Willett. New York: Oxford University Press, pp. 143–216.
Jauniaux, E, Gulbis, B, Acharya, G, et al. (1999). Maternal tobacco exposure and cotinine levels in the fetal fluids the first half of pregnancy. Obstet Gynecol 93: 25–29.
Kendrick, JS, Merritt, RK (1996). Women and smoking: an update for the 1990s. Am J Obstet Gynecol 175: 528–535.
Kendrick, JS, Zahniser, SC, Miller, N, et al. (1995). Integrating smoking cessation into routine public prenatal care: the Smoking Cessation in Pregnancy project. Am J Public Health 85: 217–222.
Kharrazi, M, Epstein, D, Hopkins, B, et al. (1999). Evaluation of four maternal smoking questions. Public Health Rep 114: 60–70.
Khoury, MJ, James, LM, Erickson, JD (1994). On the use of affected controls to address recall bias in case-control studies of birth defects. Teratology 49: 273–281.
Klemetti, A, Saxen, L (1967). Prospective versus retrospective approach in the search for environmental causes of malformations. Am J Public Health 57: 2071–2075.
Klungel, OH, de Boer, A, Paes, AHP, et al. (2000). Influence of question structure on the recall of self-reported drug use. J Clin Epidemiol 53: 273–277.
Mackenzie, SG, Lippman, A (1989). An investigation of report bias in a case-control study of pregnancy outcome. Am J Epidemiol 129: 65–75.
Mares-Perlman, JA, Klein, BE, R, et al. (1993). A diet history questionnaire ranks nutrient intakes in middle-aged and older men and women similarly to multiple food records. J Nutr 123: 489–501.
Mitchell, AA, Cottier, LB, Shapiro, S (1986). Effect of questionnaire design on recall of drug exposure in pregnancy. Am J Epidemiol 123: 670–676.
Mongelli, M, Wilcox, Gardosi, J (1996). Estimating the date of confinement: ultrasonographic biometry versus certain menstrual dates. Am J Obstet Gynecol 174: 278–281.
Moore, DA, Carpenter, TE (1999). Spatial analytical methods and geographic information systems: use in health research and epidemiology. Epidemiol Rev 21: 143–161.
Munger, RG, Romitti, PA, Daack-Hirsch, S, et al. (1996). Maternal alcohol use and risk of orofacial cleft birth defects. Teratology 54: 27–33.
O'Toole, BI, Battistutta, D, Long, A, Crouch, K (1986). A comparison of costs and data quality of three health survey methods: mail, telephone and personal home interview. Am J Epidemiol 124: 317–328.
Patterson, RE, Kristal, AR, Levy, L, et al. (1998). Validity of methods used to assess vitamin and mineral supplement use. Am J Epidemiol 148: 643–649.
Rodenbeck, SE, Sanderson, LM, Rene, A (2000). Maternal exposure to trichloroethylene in drinking water and birth-weight outcomes. Arch Environ Health 55: 188–194.
Romitti, PA, Burns, TL, Murray, JC (1997). Maternal interview reports of family history birth defects: evaluation from a population-based case-control study of orofacial clefts. Am J Med Genet 72: 422–429.
Russell, M, Martier, SS, Sokol, RJ, et al. (1996). Detecting risk drinking during pregnancy: a comparison of four screening questionnaires. Am J Public Health 86: 1435–1439.
Salvini, S, Hunter, DJ, Sampson, L, et al. (1989). Food-based validation of a dietary questionnaire: the effects week-to-week variation in food consumption. Int J Epidemiol 18: 858–867.
Sobell, J, Block, G, Koslowe, P, et al. (1989). Validation of a retrospective questionnaire assessing diet 10-15 years ago. Am J Epidemiol 130: 173–187.
Stoler, JM, Huntington, KS, Peterson, CM, et al. (1998). The prenatal detection of significant alcohol exposure with maternal blood markers. J Pediatr 133: 346–352.
Swan, SH, Shaw, GM, Schulman, J (1992). Reporting and selection bias in case-control studies of congenital malformations. Epidemiology 3: 356–363.
Thompson, WD, Walter, SD (1988). A reappraisal of the kappa coefficient. J Clin Epidemiol 41: 949–958.
Weeks, ME, Kulka, RA, Lessler, JT, Whitmore, RW (1983). Personal versus telephone surveys for collecting household health data at the local level. Am J Public Health 73: 1389–1394.
Wei, EK, Gardner, J, Field, AE, et al. (1999). Validity of a food frequency questionnaire in assessing nutrient intakes of low-income pregnant women. Matern Child Health J 3: 241–246.
Werler, MM, Hayes, C, Louik, et al. (1999). Multivitamin supplementation and risk of birth defects. Am J Epidemiol 150: 675–682.
Werler, MM, Lammer, EJ, Rosenberg, L, Mitchell, AA (1991). Maternal alcohol use in relation to selected birth defects. Am J Epidemiol 134: 691–698.
Werler, MM, Louik, C, Shapiro, S, Mitchell, AA (1996). Prepregnant weight in relation to risk of neural tube defects. JAMA 275: 1089–1092.
Werler, MM, Mitchell, AA (2000). Medication use in pregnancy in the U.S. [Abstract]. Teratology 62: 363.
Werler, MM, Pober, BR, Nelson, K, Holmes, LB (1989). Reporting accuracy among mothers of malformed and nonmalformed infants. Am J Epidemiol 129: 415–421.
Willett, W (1990). Nutritional Epidemiology. Monographs in Epidemiology and Biostatistics, Vol. 15. New York: Oxford University Press.
Willett, WC, Sampson, L, Stampfer, MJ, et al. (1985). Reproducibility and validity of a semiquantitative food frequency questionnaire. Am J Epidemiol 122: 51–65.
Witschi, JC (1990). Short-term dietary recall and recording methods. In: Nutritional epidemiology. Monographs in epidemiology and biostatistics, Vol. 15, edited by W Willett. New York: Oxford University Press, pp. 52–68.
Zhang, J, Cai, WW (1993). Association of the common cold in the first trimester of pregnancy with birth defects. Pediatrics 92: 559–563.