Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.
4 The Use of Short-Term Dietary Intake Data to Estimate Usual Dietary Intake Dietary intake of an individual is not constant from day to day but varies both in amount and in type of foods eaten and, hence, in nutrient content ( intraindividual variation). There are also variations between persons in their usual nutrient intake averaged over time (interindi- vidual variation). For North American populations, the intraindividual variation is usually as large as or greater than interindividual variations and must be taken into account in any approach to nutrient assessment. RELATIONSHIP OF DAILY DIETARY INTAKE DATA TO USUAL INTAKE: Many authors have compared the reliability of data from 1-day dietary intake records and records for longer periods (Garn et al., 1978; Marr, 1971; PeRkarinen, 1970; Young et al., 1952a), and errors in usual intake estimation due to intraindividual variation have attracted considerable inter- est. Initially, the interest of scientists was stimulated by the desire to examine biological relationships in epide- miological data sets, for example, dietary intake and serum lipid levels or energy intake and lipid levels (Beaton, 1982a: Beaton et al ., 1979: Jacobs et al ., 1979; Liu et _., 1978; Stallones, 1982 ) . At about the same time, sev- eral researchers realized that when 1-day dietary intake data are used, intraindividual intake variation results in a serious bias in regression and correlation analyses. This bias can easily lead to false conclusions about the underlying biological relationships (Beaton, 1982~; Beaton _ al., 1979; Jacobs et al., 1979; Liu et al., 1978; Sempos _ al., 1985; Stallones, 1982). The concept of measurement error and the statistical approaches for dealing with it are not new. It has been applied in other fields for many 17
18 years. However, there is new appreciation of the applica- bility of these concepts to data on dietary intake. As a result of the phenomenon of intraindividual varia- tion, when one uses a fixed cutoff point for an observed intake distribution, the number of days on which intake was recorded affects the apparent prevalence of inadequate intake (Hegsted, 1972). Figure 4-1 illustrates the impact of repeated observations on the apparent distribution of intakes and on the apparent prevalence of intakes falling below a fixed cutoff point. 30 25 J 320 > - Z As LL o 10 s "usual intakes" observed over several days SD = 12.3 9 l 1 RDA] / \\ 1-day observations mean ~ \ SD = 19.0 9 66 9 \ \ \ \ \\` l L_ 0 20 40 60 80 100 120 PROTEIN g/day FIGURE 4-~. Effect of multiple days of observation on the apparent distribution of nutrient intake. From Hegsted, 1972. Reprinted by permission of the author and publishers. Copyright Gordon and Breach Science Publishers, Inc.
19 Interest in the magnitude of this day-to-day variation in intake continued in recent years (Beaton et al., 1979, 1983; Hackett _ al., 1983; Mouser and Bebb, 1981; Hunt et_ al., 1983; Karvetti and Knuts, 1981; Liu et al., 1978; McGee _ al., 1982; Rush and Kristal, 1982; Sempos et al., 1985; Todd _ al., 1983). Earlier reports addressed the same issue with specific reference to estimating energy intake or ways to derive information about energy and lipid components (Balogh et al., 1971; Hankin et al., 1967; Kato et al. , 1973 ; Keys, 1970 ; Marr, 1971; Morris et al. , 1963 ; Tillotson et al., 1973). mese reports indicate that the magnitude of the intraindividual component, relative to the interindividual component, varies with nutrient; prob- ably with the age, sex, and sociocultural group; and with dietary methodology. This is illustrated in Table 4-1, which presents the ratio of intra- to interindividual vari- ance for various studies by nutrient and sex. Several authors noted that for nutrients with markedly skewed dis- tributions, the ratios for logarithmically transformed intake distributions were lowered (Beaton et al., 1983; Hunt _ al., 1983; Sempos et al., 1985). me table shows a consistency in the ratio of intra- to interindividual variation, in that most nutrients for all five studies produced values greater than 1. However, the data of Hackett _ al. (1983) for a group of children suggest generally higher ratios for energy, protein, total fat, and total carbohydrate. In addition, there seems to be some difference between males and females and between vari- ous nutrients. The method of assessing dietary intake may have affected the ratio of the variances, since true random error would be included in the intraindividual component. In a recent study by Sempos et al. (1985), dietary data were collected for a sample of 15 women who completed two randomly selected 1-day records per month for a total of 29 records during a period of 2 years. m e data were then analyzed for each year (Table 4-1). The various ratios were very similar in each year, suggesting that there were stable characteristics within the population. These authors demonstrated that the variance ratio does reflect the true usual intake and day-to-day variation and vali- dates other studies of shorter duration. Both Hunt _ al. (1983) and Sempos et al. (1985) noted_ _ that use of nutrient supplements altered the variance ratios. Sempos _ al. (1985) found ratios less than 1 and usually less than 0.5 for iron, thiamin, riboflavin, niacin,
20 TABLE 4-1. Observed Ratiosa of Intraindividual and Interindividual Variancesb . . _ 24-Hour 3-Day 24-Hour Recall by Record by 1-Day Recall 7-Day Recall by Young c Older d by Women e Record Pregnant Nutrient Adults Adults Year 1 Year 2 tar Menf Wameng Males: Energy 1.1 1.0 0.8 Protein 1 . 5 1.2 1.4 Carbohydrate 1.6 2.1 0.6 Fat 1.2 1.2 1.3 SFAh I.1 2.2 I.4 PUFA1 2.8 3.S 1.9 Cholesterol 3.4 5.6 1.6 Vi tamin A ~ 1.6 Vitamin C 3. 5 2.3 miamin 2.5 0.9 Ribof lavin 2.4 0.9 Niacin 1.6 2.2 Calcium 2.2 1.1 Iron 1.7 1.8 Females: . Energy 1.4 0.8 1.6 1.6 1.1 Protein 1.S 1.3 2.1 2.1 1.4 Carbohydrate 1 . 4 1.2 NRk NE 1.2 Fat 1.6 0.9 NR NR 1.2 SFAh 1.4 1.7 NR NR NE PUFAi 4.0 2.2 NR NR NR Cholesterol 4.3 4.2 NO NR NR Vitamin A 24.3 2.5 7.7 10.9 NR Vitamin C 2.0 2.8 2.3 2.5 NR Thiami n 4.4 1.6 3.3 3.9 NR Ribof lapin 2 . 2 1.8 3 . 0 3. 3 NR Niacin equivalent 4.0 2.5 NR NR NR Calcium 0. 9 1. 7 1.1 1.2 1.0 Iron 2.5 1.5 2.7 2.5 NR aA ratio of 1. 0 indicates that the ir~traindividual and interindd vidual variances are equal. A ratio greater than one indicates that intraindividual variance is greater than interindividual variance. bThe original papers contain additional data. Only those nutrient variables examined in two or more papers are included here. CFrae Beaton et al., 1979, 1983. dFrma }tunt et al . 1983 . ~ _ _ , "Fran Scopes et al ., 1985 . fFrae McGee et al., 1982. gFrom Rush et al. 1982. ~ _ , 'saturated fatty acids. polyunsaturated fatty acids. scone of the variance could be assigned to subjects. kNR=Not reported.
21 and vitamin C ( as well as for some other nutrients not included in Table 4-1) when food intakes with supplements were analyzed. PROCEDURE FOR ADJUSTING INTAKE DATA When the intraindividual variation of dietary intake and the nether of days of observation are known, it is possible to determine reliability of dietary intake data for each particular person. The usual intake for each person lies within the bounds described by the following equation 95% of the time if the day-to-day variations are normally distributed: + 2 x SD (intra)/ By, where SD (intra) is the measure of intraindividual varia- tion and n is the number of observations for the individ- ual person. The following discussion assumes normally dis- tributed (Gaussian) data, even though this is rarely the case with food intake data. The appropriate transformations to obtain Gaussian distributions are discussed later. When the number of days of observation increases from 1 to 4, the confidence limits are reduced by one-half, and the reliability of the estimate of usual dietary intake is improved accordingly. If one considers the actual magnitude of intraindividual variation, however, the results are some- what disheartening. For example, the intraindividual coef- ficient of variation for energy in adults is approximately 25% of the mean (Beaton et al., 1979). This suggests that about 3 weeks of intake data are needed to estimate usual energy intake with confidence limits of approximately +10%. When determining usual intakes in populations, there is a need not for reliable estimates of the dietary intake for each person but, rather, for reliable estimates of the dis- tribution of usual intakes for the population (Hegsted, 1972). Unlike individual intakes, the distribution of usual dietary intakes for the population can be approximated from a modest number of repetitions of the 1-day intake data; however, seasonal variations in intake and variation between weekdays and weekend days must also be taken into account in the data collection.
22 The overall variability in a distribution of dietary intake can be described in the following terms: V(total) = V(inter) + V(intra)/n, where V(total) = total variance of data (square of observed SD), V(inter) = between-subject variance, V(intra) = within- subject variance or residual error teem in an analysis of variance, and n = number of days of intake data. In this equation, the interindividual variation represents the dis- tribution of usual intake referred to previously. Statistical theory allows us to derive an estimate of the distribution of usual intakes, given the observed mean, the total variance, and an estimate of the intraindividual vari- ance. Replicated observations of 1-day dietary intake are needed to obtain an estimate of intraindividual variation. In theory, the replicated observations should be independent of one another in tome rather than on consecutive days, although it is not yet known whether this is important in practice. The magnitude of the intraindiv~dual variation can be estimated by analysis of variance (ANOVA). If the original distribution is not normal, the distribution must be trans- formed into a more normal form before the ANOVA procedure is applied. (See Appendix A for an example of a logarithmic transformation for this purpose.) To adjust the intake distribution, the deviation of each point from the popula- tion mean is multiplied by the ratio of the interindividual standard deviation to observed standard deviation. With this nonpar~etric procedure, it is not necessary to assume a perfect fit of the normal distribution, and the distinc- tive shape of the original distribution is preserved. The adjusted data are then transformed back to the original scale for subsequent analyses. Figures 4-2 and 4-3 depict two distributions derived from Nationwide Food Consumption Survey (NFCS) data that have been adjusted to estimate the distribution of usual dietary intake of protein for males and iron for females. Appendix A provides full details of the approach used in generating distributions for this report. Although ANOVA is not recommended as the only approach to estimating inter- individual variation and eliminating the effects of day-to- day variability, it can serve as an example of possible
23 14.8 11.9 of o ~ C 8.9 o ~ 5.9 at 3.0 o A ,W, 0 44 / \ USUAL INTAKE ken\ \ \~ 1-DAY INTAKE - 88 132 176 220 PROTEI N INTAKE, g/day FIGURE 4-2. Comparison of 1-day and adjusted distributions for protein intake by male adults. Derived from the 1977-1978 NFCS data analysis described in Appendix A. Q Q - 7.1 o ~ 5.3 J 2 3.5 ~9 t.8 o O | ~ USUAL INTAKE ted/) \ '1 Am\ 1 -DAY I NTAKE g - __ 6.3 12.6 ~ 8.9 25.2 31.5 I RON I NTAK E, mg/day FIGURE 4-3. Comparison of 1-day and adjusted distributions for iron intake by female adults. Derived from the 1977-1978 NFCS data analysis described in Appendix A.
24 approaches (Trumpler and Weaver, 1953). The method of normalizing the original distribution should be appropriate to the data set under study (Box and Cox, 1964). The current NFCS is designed to collect information for 3 consecutive days for all subj eats . There is a possibility that dietary intake on consecutive days is correlated within the individual. If this is true, the statistical power of the estimates is reduced. On the other hand, the subcom- mittee believes that 3 days of observation may be more than is required for the derivation of the distribution of usual intakes. For purposes other than the analysis of dietary adequacy, such as using dietary intake data for multivariate analysis, data for additional days would probably be re- quired. All goals of the survey must be considered when the final decision Is made. In summary, the Impact of day-to-day variation in intake contributes to errors in estimations of usual intake. If the survey data include an adequate number of independent replicate observations of intake measurements, methods can be used to adjust the observed intake distribution to gen- erate a good estimate of the distribution of usual intakes. In a large survey, such as the NFCS, this approach is feast ble and may even permit collection of fewer data than are now collected, given an appropriate sampling design. The feasibility of this reduction depends, of course, on other uses of the data. For example, if providing descriptive data on patterns of food use other than intake is a purpose of the survey, more replications may be needed. There is no need to collect more days of dietary data per individual than in the recent NFCS to implement the analytical approach described in this chapter for adjustment of the intake distributions. The only additional cost involves the statistical resources needed to design and analyze the data. -