Click for next page ( 49


The National Academies | 500 Fifth St. N.W. | Washington, D.C. 20001
Copyright © National Academy of Sciences. All rights reserved.
Terms of Use and Privacy Statement



Below are the first 10 and last 10 pages of uncorrected machine-read text (when available) of this chapter, followed by the top 30 algorithmically extracted key phrases from the chapter as a whole.
Intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text on the opening pages of each chapter. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

Do not use for reproduction, copying, pasting, or reading; exclusively for search engines.

OCR for page 48
7 Errors in Nutrient Intake Measurement All measurements have components of random errors and systematic bias. Dietary intake measurements are no excep- tion. In developing an approach to the analysis of dietary data, it is essential to consider the effect of both types of error. The estimates of nutrient intake are based on data from dietary surveys, data in food composition tables, and the computation of the nutrient intake from these two data sets. Each data set has sources of random error and bias; there is also potential for bias in the computation process. Thus, the subcommittee discussed sources of bias and variability in the data on dietary intake, food compo- sition, and computation of nutrient intake. This chapter presents the results of its analysis of the impact of random error and systematic bias on the estimates of the prevalence of inadequate intake. SAMPLING VARIATION Random Error By chance, the persons randomly selected in dietary surveys may not be representative of the reference popu- lation. Using statistical theory, one can derive minimum estimates of this sampling error, which would be increased by other sources of variability such as those discussed in this chapter. Systematic Bias Systematic bias in sampling can also occur. For example, in such surveys as the Nationwide Food Consumption Survey (NFCS), the sample design is based on households. There- 48

OCR for page 48
49 fore, the homeless are systematically unde~represented. Self-selection (i.e., survey participation by consent) can also produce an unrepresentative sample of the U.S. popu- lation and result in systematic bias. When there are sev- eral components to the collection of data, fewer persons may respond in one component than in another. For example, fewer persons may complete dietary diaries than respond to a household interview. This bias would be especially pro- blematic if those who completed diaries were better orga- nized and better educated than those who were only inter- viewed. The magnitude of bias from sample design and from nonresponse requires special research not usually included in surveys of this kind. At a minimum, it is important to determine the probable direction of any such bias and, where possible, it is preferable to estimate the magnitude as well. ERRORS IN ESTIMATING USUAL NUTRIENT rNTAKE Errors in Reporting Usual Food Intake Day-to-Day Variation in Intake. The intake of concern is the average intake of individuals across time. This usual intake is believed to af feet tissue levels of the nutrients and body functions. Intake on a particular day does not reliably portray the usual intake. The Impact of day-to-day variation in intake is dis- cussed in some detail in Chapter 4, where the subcommittee discusses a method to adjust dietary intake to control for day-to-day variation. Although several statistical tech- niques might be used for this purpose, the subcommittee chose to use an analysis of variance procedure for the analyses presented in this report. Variability In Reporting and Recording Food Intake. A . close examination of errors in reporting, recording, and coding data from food intake surveys is helpful in ident~- fying errors that are random within a person and those that are systematic. Random error. A respondent may sometimes overreport and sometimes underreport in a random fashion. Variation also occurs between persons when one person underreports and another person overreports. Even if these errors are random, they present a problem in the analysis of popu-

OCR for page 48
50 ration data. However, systematic bias involving entire subgroups of the population may have more serious impli- cations for the analysis of such data. For example, if all or most members of the lower socioeconomic groups overestimate intake of such status foods as meat, biased estimates of iron intake may result. Systematic bias. In the NFCS, the importance of systematic underreporting in the total population has been recognized. As a result, the U.S. Department of Agriculture (USDA) has undertaken or funded numerous studies that address these problems. The literature relating to these questions is reviewed briefly in the following paragraphs. The goal of any dietary survey is to measure what people eat-more precisely, to measure what foods and supplements people habitually eat or usually ingest. Dietary surrey methods can be classified into two general categories: methods bared on memory (recall) and methods based on records. In memory-based procedures, subjects are asked to recall all food and drink consumed over a 'specified time, usually 24 hours. Record-keeping methods take Several forms: a common procedure is to ark subjects to keep a diary recording all foods and beverages consumed during a Specified nether of days, most commonly for per- iods of up to 7 days. In some cases, respondents are provided with plastic or paper food models and measuring devices to aid in estimating portion sizes. In other cases, they receive only specific written or oral instruc- tions or both. As mentioned earlier, the recall and record methods were combined in the collection of 3-day dietary intake data for individuals in the 1977-1978 NFCS. In recent years there have been several reviews of dietary methods (Bear and Laus, 1982; Becker et al., 1960; Burk and Pao, 1976; Marr, 1971) as well as many papers concerned with the quality of specific methods. The aspect of quality of greatest concern is validity, especially with regard to systematic bias in data on dietary intake. A valid method is one that measures what it intends to mea';ure--the true intake of subjects. Dietary methods can be validated only by knowledge of tore intake or by some Sensitive laboratory measurement of intake. Block (1982) notes that it is difficult to ascertain true

OCR for page 48
51 intake over long periods. Consequently, validation fre- quently consists of comparing one dietary survey method with another. Such validation is limited by the lack of agreement on a "true" or reference method. Similarities in the two methods may fail to detect common errors when the two methods are compared. Failure to record accu- rately all foods eaten during a survey may influence sub- sequent recall. It may also sensitize the respondent's power of observation, so that a subsequent recall is more complete than the initial daily record, as in a survey or nutritional assessment. Unobtrusive observations of actual intake have been made of people in various institutional settings, such as children in grade schools (Comstock et al., 1981; Graves and Shannon, 1983; Lachance, 1976; Meredith et al., 1951), students in college cafeterias (Krantzler et al., 1982; Mullen _ al., 1984; Raker, 1979), elderly people at con- gregate meal settings (Gersovitz et al., 1978; Madden et al., 1976), patients in nursing homes (Caliendo, 1981), adolescents in a metabolic research unit (Greger and Etnyre, 1978), young boys consuming food at home for a 1-day period (Stunkard and Waxman, 1981), lactating women confined to a hospital (Linusson et al., 1975), and massively obese patients in a clinical research center (Bray et al., 1978). Settings such as these are optimal for these studies because portions served are standard- ized or predetermined. These observations have usually been made for short periods or even for single meals, although observations as long as 28 consecutive days were reported by Mullen et al. (1984) and Krantzler et al. (1982) in their study of students in a dormitory dining hall. Institutional routines may, however, alter the eating patterns and recall as compared to the general population. Caveats about validation must be taken into account before the accuracy and systematic bias in food intake data are addressed. For example, are there systematic errors in the dietary intakes reported by some subgroups of the population, such as young and old subjects, or in the reporting of particular foods or food groups--biases that would influence the reported nutrient intakes? the results of several validation studies indicate that low intakes may be overreported and high intakes under- reported, with a resultant flat-slope syndrome (Gersovitz

OCR for page 48
52 _ al., 1978; Madden et al., 1976; Stunkard and Waxman, 1981), although mean intakes for groups may be accurately recalled. However, the flat slope syndrome may be nothing more than the attenuation of the "lope that arises from random error in the independent variable, as has been described by many authors. Systematic biases may exist in dietary reporting by selected population subgroups but have not been suffi- ciently addressed to warrant conclusions. For example, Campbell and Dodds (1967) used a 24-hour recall to collect dietary information from elderly and young patients hospi- talized for various lung disorders. Their purpose was to test the extent of error when a 24-hour recall was used to collect information from elderly subjects whose failing memory may affect recall. The 24-hour recall data were checked by prompting subjects based on a known menu. In this study, the elderly subjects underreported calories more often than did younger subjects, and men underre- ported more calories than did women. Dietary intakes obtained without probing were approximately 259 lower than those obtained with probing. Marr (1971) concluded that it is difficult to obtain valid dietary data from young children. For example, it is necessary to rely on a surrogate, such as a parent, to provide information about the intake of young children. In addition, caretakers other than parents frequently provide food for the child, making it difficult to secure the necessary data. In the United States, food is increas- ingly eaten away from home by young children as more and more preschoolers attend childcare facilities. A preliminary study of 29 children of preschool age was designed to examine the relative usefulness of the 1-day food record and the 24-hour recall ( H. Smiciklas-Wright and P. Hol mberg, Pennsylvania State University, personal communication, 1985 ) . The investigators observed actual intake at a lurch consumed by children attending a day- care center and then asked 14 mothers to provide data on dietary intake by a 24-hour recall. Seven of the 14 moth- ers provided no data for the meals consumed while their children were at the center. The remaining mothers failed to identify menu items correctly, overestimated portion sizes, or bath. The 15 parents who provided information by a l~day record did provide more complete data, although these parents also tended to m' sidentify foods or over-

OCR for page 48
l 53 estimate portion sizes. For data interpretation it is important to know whether a surrogate reporter was pres- ent at all meals. Retrospective diet histories have been compared with direct measurements of food intake by obese patients hospitalized in a metabolic unit (Bray et al., 1978). Three retrospective assessments of diet history were obtained at Month intervals. In the first history, energy intake was underestimated. By the third one, esti- mates of energy intake had risen--particularly because of improved reporting of alcohol intake. Thus, the third history probably provided the best correlation between true and reported energy intake. Lansky and Brewnell (1982) also examined the accuracy of food records of applicants in a behavioral weight- reduction program. Thirty women estimated the quantity of 10 foods that were displayed in small and large con- tainers. Container size did not influence quantity estimates, but the quantities of all 10 foods were over- estimated. The amount of overestimation varied from a small overestimation for cola and orange juice to a large overestimation for potato chips, ham, and turkey. Neither Brav et al. (1978) nor Lansky and Brownell ( 1982 ) reported comparable data on subjects of normal weight to determine whether the errors in reporting intake and recording serv- ing sizes were restricted to obese persons. A. Blake (Pennsylvania State University, personal communication, 1985 ) found that serving sizes were underestimated by both obese subjects and persons within normal weight ranges. ThiS investigator provided preweighed portions of foods at a lunch attended by subjects who were contacted the next day for a 24-hour recall. The majority of subjects, both obese and normal, underestimated the servings they had eaten. Sopko _ al. (1984) found that obese males reported from 78% to 93% of their actual intake in a controlled feeding experiment. Although the study was not a true validation, Hallfrisch _ al. (1982) observed that on a 7-day diet record, men reported 80% and women reported 65% of the con- sumed calories in subsequent 6-week experimental periods. These results call into question the validity of recall or recording of food intake. Reports of food and energy intakes only slightly higher than basal energy requi re-

OCR for page 48
54 meets of adult women are difficult to reconcile with popu- lation data showing a 40% prevalence of overweight and a 25.7% prevalence of obesity, as indicated by body mass index (Van Itallie and Woteki, 1985). Selective underreporting is also suggested by compari- sons of beer, wine, and distilled spirit consumption, as reported for adult males and females in the NFCS and National Health and Nutrition Examination Survey (NHANES), to estimates of alcohol disappearance, as reported by the Bureau of Alcohol, Firearms, and Tobacco Control.(Pao et al., 1982). Similarly, Schnakenberg et al. (1981), studying the reported food intake of 62 men in a military dining hall for a 7-day period in comparison to observed and weighed intakes, found that 13% underestimated their caloric intake by more than 30% and 34% underestimated their caloric intake by 10% to 20~. Only 5% overestimated by more than 30~. In contrast, de St. Jeor (1980), who monitored weight status to validate the accuracy of energy intake reported by paid, educated subjects over a 12-week period, concluded that recorded intakes were an acceptable measure of actual intake. As reported earlier, A. Blake (Pennsylvania State University, personal communication, 1985), comparing the recalled intake of meals to observed and weighed intakes by obese and nonobese subjects, found no difference between the two groups, but did confine a tendency to overreport small intakes and underreport higher intakes. She found estimated portion sizes to.vary from actual sizes by less than 10% for most food items. One reason for the discrepancy between.actual and reported intake may be the difficulty that respondents have in estimating portion sizes. Guthrie (1984) found that for 13 food items, from 6% to 75% of adult male and female subjects reported portion sizes that varied by more than 50% from the weighed portion sizes. However, only 26% of the respondents consistently under- or overrepor~ted al1 items in the meal. H. Smiciklas-Wright and H. Guthrie (Pennsylvania State University, personal communication, 1984) showed that the ability of college-age students to estimate within 1 oz the volume of fluid contained in drinking glasses varied from 30% to 70%, depending on the size and shape of the container. Young et _ . (1952) found that for a variety of foods the direction of error was generally in overestimating por-

OCR for page 48
55 tions. Compensations were made for this bias by using zero for all omissions of data when calculating means. They con- cluded that for the group, errors in the estimation of por- tion sizes for most food types are probably within 20% of actual portions, except for children's recall, which has a greater error. Krantzler et al. (1982) reported on 24-hour recall and 3- day or 7-day records of students who were observed periodi- cally over a 28-day period during which they took their meals in a dormitory dining hall. They reported that foods eaten regularly, i.e., those contributing the major part of a meal, were better reported than such foods as condiments, nuts, and seeds. Estimates of dairy products, meat, and fish were the most accurate. Guthrie (1984) reported that few subjects in her study forgot main meal items, but one in six respondents forgot to mention salad dressing. Because two-thirds of the respondents used more than 20 g (about 130 kcal) of salad dressing, that omission alone could represent a sizable error in daily intake. The foregoing discussion suggests that errors in reporting dietary intake undoubtedly exist. The studies suggest also that the direction of errors varies from study to study and perhaps from population group to popu- lation group. In a separate exercise, the committee cam- pared the distributions of intakes reported in several recent large surveys. There were differences, suggesting bias in either estimation of food intake or food composi- tion data, but these were not consistent across nutrients. The evidence is not sufficiently consistent to suggest how these systematic biases affect estimates of individual nutrients. It is clear, however, that they can do so markedly. Because the 1977-1978 NFCS did not include data on nutrient supplements, it certainly underestimated the intake. Variability Due to Coding and Analysis of the Nutrient Content of Foods Sources of Technical Errors in Food Composition. The technical errors in food composition data that influence the interpretation of food consumption surveys fall into three broad categories: true random variability of the composition of individual food items, biased food com- position data, and the differences in bioavailability of individual dietary nutrients.

OCR for page 48
56 True random variability. Variations in the com- position of a food item from the population mean as given in their classification codes reflects the true variation in nutrient content of foods due to differences in produc- tion practices and to the effects on raw materials exerted by such variables as soil, fertilizer application, weather, pest control, and genetic variation. Postharvest physio- logical changes, storage, and processing also contribute to the true variability of the nutrient content of individ- ual foods. Estimates of this normal variation were included in the current USDA food composition data tables, which were used for the sensitivity analysis of dietary intake information included later in this chapter. This analysis indicates that normal tood composition variation contributes a significant portion of the total variation in the estimated intake of given nutrients. However, the sub- committee suspects that an all-out effort to decrease the standard error of the mean of the data in the f cod composition data base is not needed in most cases. Even though food composition contributes to variation, coeffi- cients of variation due to errors in the computation of intakes for most nutrients are generally Small and the overall error is modest. The subcommittee suggests that sensitivity analysis be applied to the specific nutrients to determine the effect of reducing the error of the mean. The results of the sensitivity analysis should then be used to develop priorities designating which nutrients require further refinements either by assaying more sam- ples or developing more precise analytical methodology. Biased food composition data. Errors in food composition Data result wnen one aaCa are cons~snen='y incorrect due to the methods of data collection or an~l- ysis. Often neither the direction nor extent of bias is known for individual foods. Common causes of biased data are incorrect identification of the food its, the use of inappropriate analytical methodology, and the use of imputed values. Incorrect identification of the food being anal- Yzed, despite the precision of the analytical technique used in the assay, leads to biased data. Thus, the subcom- mittee encourages the USDA to continue with its ongoing efforts to improve the food nomenclature system. The nutrient values in the current food composition data bases for food categories are weighted averages of the con-

OCR for page 48
~7 tents of the individual foods that fall within the category of the listing. Such data are acceptable for calculating nutrient intake if the nutrient content of each food is ran- domly distributed about the mean for the food category. In such cases the estimated nutrient intakes can be reasonably calculated from the mean content of the food category. How- ever, the use of such mean values to estimate nutrient in- take can lead to significant bias when the nutrient content of the food item is not randomly distributed about the mean of the food group. For some categories of processed and manufactured foods, means are not representative of values for certain brands because their nutrient contents are controlled by recipes or processes with unique formulations significantly dif- ferent from the category mean. Formulations vary to this extent for only some foods and even then for only some nutrients in those foods. For example, wide variations in nutrient content could be found in vit~min-fortified fruit juices and breakfast cereals, the sugar content of break- fast cereals, the fat content of bakery products, and the sodium content of many different prepared foods and meals. When brand loyalty is considerable, that is, when consumers consistently use the same brand rather than randomly selecting similar products, and when the mean nutrient con- tent of the brand item differs from that of the category mean, biases will occur in the calculated estimations of nutrient intake. Such problems can be corrected by using brand-specific composition data. Since the addition of brand name identifiers may increase the complexity, and therefore the cost, of the food composition data and food consumption surveys, the subcommittee suggests that sensi- tivity analysis be used to determine which brands need to be identified by brand name to improve estimates of the intake of particular nutrients. The analytical methods constitute a recognized source of bias in food composition data. In some cases, the methods do loot measure all chemical forms of the nutrient, thus leading to underreporting for foods that contain the unmea- sured chemical forms. Other nutrient assays are inhibited by some food components, and when such components are pres- ent, the nutrient content will be underreported. Certain assays respond to components other than the nutrient of interest, and when such substances are present, the nutri- ent content will be overreported.

OCR for page 48
~8 Where there are no direct data on the nutrient composi- tion of a particular food, nutrient values can be calcu- lated by sting the contributions of each component of the formulation, if the food item is formulated from several other food components with known values. Although the resulting figures usually give acceptable estimates of the nutrient content of the foods, in some cases a value is imputed for the nutrient by using the nutrient content of a similar food and bias may be introduced. For example, lacking analytical values for corn syrup, known composi- tion values for molasses might be imputed under the assump- tion that corn syrup had the same composition. This results in an overestimation of iron intake, because molas- ses contains a good deal of iron and commercial syrup has none. Imputed and calculated values in the current USDA food composition tables can be identified because the entry for the number of samples is left blank. However, there is no way for the users of the tables to determine directly if the other nutrient values are imputed or calculated. The subcommittee recommends that the practice of imputing food composition values be avoided. When sen- sitivity analysis shows that imputing the content of cer- tain foods may result in significant errors in estimating the prevalence of inadequate intake, the components of those foods must be analyzed. ~~ However, no person's dietary information should be omitted because there are no analyzed data for converting foods to nutrients. Thus, values for many foods must be imputed. It is important therefore that these imputations be reasonable estimates. For instance, assigning zero as an imputed value simply because the data are not available is not a reasonable estimate. If imputed values are used, they should be flagged so that their impact on the data analysis can be considered. Biases due to food composition data--incorrect assessment of differences in bioavailab~ bed ~ u.r=~. The selection of criteria for assessing nutritional intake is complicated by the variation in absorption of nutri- ents. For example, evidence indicates that the bioavaila- bility of iron varies with its chemical form in the diet (e.g., heme versus nonheme; reduced iron versus iron phos- phates); the presence or absence of absorption enhancers in the meal (e.g., ascorbic acid or the "meat factor"); the presence or absence of compounds in the mea'-, such as phytates and tannic acid, which reduce the absorption of iron; and the physiological status of the small intestine

OCR for page 48
59 - - as it relates to the state of the iron stores of the indi- vidual (Morris, 1983). The assessment of the iron intake of individuals has several components: the iron require- ment of different age and sex groups, the food composition data, and often an assessment of the bioavailability of iron in the meals. To take into account the bioavailabil- ity of iron in the diet in the 1980 RDAs, a factor of 10 was applied to the actual requirement to account for absorp- tion losses. In contrast, data based on food composition do not usually contain any direct information on the bio- availability of iron in individual foods. However, this is not always the case. For vitamin E, most food composition data bases do have a built-in estimate of bioavailability for each food. The subcommittee believes that inclusion of the bioavailability of iron in calculation of requirements introduces a potential source of bias in the assessment of nutrient intake. Some forms of iron are available at much less than the 10% level and some at higher levels. A1- though the use of a fixed level of absorption does not take into account the influence of the iron sources on its bio- availability, a 20% level has been suggested as the upper limit by an FAD/WHO (1970) committee and some of the cal- culetions in this report are based on this limit (see Appendix B). It is possible to improve the data on bioavailability by beginning with the actual requirement for absorbed iron for each age and sex category. This can be done by removing the correction factor for the lack of iron availability from the stated iron requirement). The concentrations of each chem- ical form of the iron in each food in the food composition data base can then be listed together with the concentra- tions of each iron absorption enhancer and inhibitor for each food in the data base. In this way it is possible to develop algorithms to calculate the biologically available iron for each meal based on the chemical form(s) of the enhancers and the inhibitors in the meal. Although this approach will provide a better estimate of the iron intake of individuals and thus populations, it requires the devel- opment of new computer algorithms. It thus requires more research, which must be guided by sensitivity analysis to avoid research and analysis of marginal utility. Nutrient Data with Probable bias In a recent report of the National Research Council, the current status of the methodology for nutrient analysis in

OCR for page 48
60 foods was discussed (NECK 1984). Methods for assaying sev- eral of the nutrients of concern to this subcommittee were found to be less than adequate. In particular, the authors found that the methods for vitamin A, carotenoids, vit ~ n B12, vitamin C, and folacin were such that there was only a fair probability that these methods would produce correct values. Following are detailed discussions of the status of these nutrients. Vitamin A. Data on the composition of foods containing preformed vitamin A (whether naturally present or added during processing) appear to be reasonable, since the methods used to obtain them appear to be reasonable assay systems. Thus, the subcommittee believes that estimates of the dietary intake of vitamin A from these data are reasonable. Carotenoids. Approximately 50 carotenoids possess vitamin A activity, each apparently with its own biological potency. Currently the vitamin A contributions of all these compounds are lumped together in food composition data bases as retinal equivalents. The chemical assay of the carotenoids is very complex (particularly in fruits and vegetables), and the current techniques used for food assay are not adequate for the determination of all the carotenoids in foods. Further- more, there is no agreement on the assignment of biological potency as vitamin A for each carokenoid isomer. The subcom- mittee suggests that studies be undertaken to determine the concentration of each carotenoid isomer in fruits and vegeta- bles. If such studies are undertaken, consideration should also be given to the concept of measuring all carotenoid isomers because of the interest in the apparent protective effects these may have as anticancer agents. There is no apparent reason to believe that only the carotenoids with vitamin A activity may have anticancer activity. Vitamin Big. The current acceptable assay for vitamin . . . . . . . . _ B12 is the classical microbiological assay using Lacto- bacillus. The technique appears to work well with raw com- .. modifies but not for processed foods that contain microbial growth inhibitors. The recently introduced protein-binding assays have been shown to respond to B12 isomers, which have no vitamin activity. Data on the correct vitamin B12 content of foods are therefore available for only a few foods, but sensitivity analysis will probably show that better, more complete analyses are only necessary for a few other foods.

OCR for page 48
61 Vitamin C. Both ascorbic acid and dehydroascorbic acid have vitamin C activity. There are several acceptable methods for measuring both forms. These appear to be ade- quate for assaying vitamin C in fresh products. Vit~Tn~n C is very labile to air oxidation, and the use of improper sample preparation can lead to an underestimation of the vitamin C content of fresh products. In the case of processed foods, the situation is more difficult. Almost all the methods that measure the two forms of vitamin C also detect isoascorbic acid, and because isoascorbic acid is commonly used as an antioxidant in the food-processing industry, but has no vitamin C activity in humans, assays for the vitamin C content of processed foods have the potential for overestimating the vitamin C content. Folacin. The class of chemical compounds named folacin consists of a complex mixture of chemical isomers with various oxidation states and different numbers of glutamate residues. There is no single chemical, biochemical, or microbiological assay that will accurately measure all the forms of this nutrient so they can be related to folacin activity in humans. Furthermore, the subcommittee does not know of any acceptable combination of techniques for mea- suring all the Homeric compounds with folacin activity. The standard assay is a microbiological one that has differential responses to different isomers. It is not known whether or not humans have the same response to the individual folacin isomers as does Lactobacillus casei; however, there is no reason to assume that they do. It is known that there are a number of additional compounds that interfere with the existing assays or that alter the bio- availability of each folacin isomer. The interactions of these assay inhibitors and of compounds that alter the bioavailability are not well understood nor, apparently, have all the interfering compounds been identified. An accurate, sensitive chemical assay for all the folacin isomers is needed to permit the resolution of the complex problems associated with assaying the folacin content of foods. riven the difficulties with assays of the folacin isomers, the data on the folacin content of food items do not appear to be reliable and the subcommittee does not believe that accurate estimates of the dietary intake of folacin can be obtained from the current food composition data bases.

OCR for page 48
62 IMPACT OF SYSTEMATIC BLAS The preceding discussion has suggested that sources of random variation in either the food intake estimate or the food composition data affect confidence in the estimates of the prevalence of inadequate intake. For food intake data, the standard error of the composition data for individual foods may result in a relatively small under- or overesti- mation of prevalence. The models do not include systematic bias, such as that derived from consistent under- or over- estimation of either total nutrient intake or the nutrient content of foods. The following paragraphs address the impact of that type of effect. As noted, there are no valid estimates of the magnitude (if any) of bias in estimates of food intake from the NFCS data. The subcommittee emphasizes the need to determine if there is such a bias and if so, its extent so that the methods of estimation can be improved. An analagous situ- ation holds for food composition data. In one study, Wolf (1981) estimated nutrient intakes from food records, from the 1963 USDA food composition data base (USDA, 1963), and from direct chemical analysis (see Table 7-1). The analyses suggest that there was a definite bias toward overestimation of true iron intake. The low regression slope and correlation coefficient suggest that the bias may apply to only some classes of foods rather than to all foods. That is, there may have been a methodological error in the food composition table for some classes of foods. USDA's food composition tables have been revised since those used by Wolf (1981). Major changes were made in the iron data for some classes of foods. At the request of the subcommittee, Wolf and his colleagues have recalculated the data from the original study by using the new food composi- tion tables. Table 7-1 indicates that there was a systematic bias in the earlier food composition tables, a bias that has now been removed for iron. The comparison illustrates also that bias in composition data must be investigated separately for each nutrient in each food composition table instead of being regarded as generically applicable. Marr (19713 has presented descriptions of several compar- isons of calculated and measured intake; the direction of bias is inconsistent across studies and across nutri-

OCR for page 48
63 TABLE 7-1. Comparison of Calculated and Measured Food Intakes Using Revised Flood Composition Tables for 22 Subj ectsa Estimated Nutrient Intake Correla- tion (R Group Mean (mg/day square) ~ l~5l=~37URbii67~ic Regression Nutrient Equation Calcium Y = 0.82X + 71 0.89 832 762 Iron original Y = 0.49X + 4.3 0.25 12.8 10.6 recalcu- lation Y = 0.83X + 2.09 0.33 lO.S 10.6 Zinc Y = 0.68X + 3.5 0.43 7.4 8.6 Copper Y = 0.lSX + 0.88 0.02 0.89 1.01 Potassium Y = 0.82X + 0.26 0.72 2,680 2,370 Sodium Y = 0.35X + 2.2 0.19 2,800 3,190 aFrom Wolf, 1981. bSelf-administered intake record. CChemically measured content of duplicate meal .. ents. Depending on the study design, such comparisons may be affected by variation and bias either in the food composition table alone or in combination with the food intake record. When interpreting the results of such studies, one must recognize that the composition of individual food items vary and that the food composition table, at best, presents the average composition of a class of foods. That is, one should not expect perfect agreement between calculated and measured composition for an item of food or for a diet. This effect is demonstrated in Appendix E. Moreover, the introduction of this variation in composi- tion (or, if one wishes, random variation) into the esti- mate of food composition in comparison to the true com- position of the foods consumed will affect the slope of regression analyses in which calculated and directly mea- sured composition are compared. To some degree, the flat slope syndrome described above and in the nutrition liter-

OCR for page 48
64 ature (e.g., Marr, 1971) can be attributed to this type of effect. (A similar phenomenon will occur if, for example, 1-day intakes are compared with multiple day intakes: the intraindividual variations in the data will be different, and this may affect the regression slope while leaving group mean comparisons unaffected.) Although there are no direct estimates of the magnitude of bias that may exist in computed nutrient intakes reported in recent NFCS or other large-scale survey data bases, it is possible to demonstrate the potential effect of such bias on estimates of the preva- lence of inadequate intake. This is illustrated for protein in Table 7-2. Here the observed distribution of intakes has been systematically increased or decreased by TABLE 7-2. Potential Impact of Systematic Bias in Either Food Composition Tables or Food Estimates on Estimates of the Prevalence of Inadequate Intakesa . Adjustment to Intake Distribution (Mock Systematic Bias in Intake Estimate) : Group Mean Intake Estimated Bias in (g/day) Prevalence (%) Estimate (%) Original ~ 20 g 111.2 0. 1 - 2 .1 Original + 15 g 106.2 0.3 - 1.9 Original ~ 10 g 101.2 0.7 - 1.5 Original + 5 g 96.2 1.2 - 1. 0 Original 91. 2 2.2 0 Original - 5 g 86.2 3.6 + 1.4 Original - 10 g 81.2 5.6 ~ 3.4 Original - 15 g 76.2 8.4 + 6.2 Original - 20 g 71. 2 12 .3 +10. 1 . aData from 1977-1978 NFCS for protein intake by adult males. adding a constant amount of protein to each intake. This process shifts the total intake distribution upward or downward but does not change the shape of the distribu- tion.

OCR for page 48
65 The magnitude of the effect on estimates of the preva- lence of inadequate intake will depend on the relative positions of the distributions of intake and of require- ment. Nevertheless, it is readily apparent that for any such estimate of adequacy, systematic bias in either food intake estimates or food composition data will result in erroneous estimates of prevalence. This effect is not specific to the probability approach. It would also occur if fixed cutoff points had been used.