ARTICLE IN PRESS
Environmental Research 103 (2007) 325–330 www.elsevier.com/locate/envres
Estimating dietary consumption patterns among children: A comparison between cross-sectional and longitudinal study designs Marjory L. Givens, Chensheng Lu, Scott M. Bartell, Melanie A. Pearson Department of Environmental and Occupational Health, Rollins School of Public Health, Emory University, 1518 Clifton Road, NE, Atlanta, GA 30322, USA Received 13 March 2006; received in revised form 23 June 2006; accepted 6 July 2006 Available online 14 August 2006
Abstract Estimating dietary intake for children is an essential component of conducting pesticide exposure assessments given the fact that children are predominantly exposed to certain pesticides, such as organophosphorus pesticide, through dietary intake. Different study designs and their respective sampling methodology utilized to estimate food consumption patterns can signiﬁcantly alter the parameter estimates and the variability in the values obtained. This study investigated the impacts of study design on overall estimates of dietary intake by applying the temporal sampling characteristics used in cross-sectional approaches, as in The Continuing Survey of Food for Intakes by Individuals (CSFII), to food consumption data collected in a longitudinal manner via a bootstrap sampling technique. We examined the precision of time-averaged dietary intake estimates under various sampling schemes and explored the contribution of seasonality toward the dietary patterns. A comparison between the estimates of food consumption obtained from the bootstrap replicates and the longitudinal study estimates indicate that variability is signiﬁcantly decreased when employing a longitudinal study design. Moreover, both between and within-subject variability decreases when individuals are followed over an increasing number of days. Finally, within the longitudinal study cohort, we observed a seasonal component to dietary intake for fruits and grains. Our ﬁndings suggest that longitudinal dietary surveys offer substantial improvements for exposure assessment compared to a standard cross-sectional design. r 2006 Elsevier Inc. All rights reserved. Keywords: Children’s pesticide exposure; Dietary intake patterns; Dietary pesticide exposure; Organophosphorus pesticides; CSFII
1. Introduction Biological consequences of pesticide exposure have been reported for incidences of heightened exposure (Dharmani and Jaga, 2005), but little is known about the effects of chronic, low-level pesticide exposure, particularly for children. Several epidemiological studies have observed an association between pesticide exposure and cancer or weakened immune systems in children (Repetto and Baliga, 1996; Daniels et al., 1997; Zahm and Ward, 1998). In addition, the risk of developing cancer later in life is increased for individuals exposed to carcinogens at younger ages (Zahm and Ward, 1998). Consequently, it is imperative to investigate the risk of exposure to pesticides among children in order to understand the potential Corresponding author. Fax: +1 404 727 8744.
E-mail address: [email protected]
(C. Lu). 0013-9351/$ - see front matter r 2006 Elsevier Inc. All rights reserved. doi:10.1016/j.envres.2006.07.003
repercussions of chronic pesticide use in agricultural and residential settings. Of several pathways of exposure, dietary intake has been noted as a potential route for pesticide exposure (Curl et al., 2003) and very recently identiﬁed as a major source, particularly for organophosphorus (OP) pesticides (Lu et al., 2006). For children, this is especially concerning due to their eating habits, which vary both quantitatively and qualitatively from those of adults. Not only do children consume more food per body weight than do adults but their dietary preferences also drastically differ from adults (NRC, 1993). It is common for children to consume a limited menu of foods such as fruits (apples, peaches) and grain products (crackers, bread, cereal) that frequently contain pesticide residues (Fenske et al., 2003; USDA, 2005a). Therefore, children are likely exposed to disproportionately higher amounts of pesticides through dietary intake compared to adults. In fact, the National Research
ARTICLE IN PRESS 326
M.L. Givens et al. / Environmental Research 103 (2007) 325–330
Council has recognized that differences in food consumption and therefore in dietary exposure to pesticides account for much of the dissimilarity in pesticide-related health risks that were found to exist between children and adults, indicating that a better understanding of dietary exposure may outweigh a focus on age-speciﬁc differences in toxicological vulnerability (NRC, 1993). To date, a prevailing methodology for assessing dietary intake is through The Continuing Survey of Food for Intakes by Individuals (CSFII), a national survey administered over the years of 1994–1996 and 1998, currently incorporated into the National Health and Nutrition Examination Survey (NHANES). This survey randomly recruited over 20,000 participants across all ages, nationwide. In response to the Food Quality Protection Act of 1996, the 1998 CSFII survey over sampled children. Thus, across the 3-year sampling period, dietary intake data was recorded for almost 10,000 children ranging from birth to 9 years of age. CSFII used a cross-sectional approach to characterize food intake for a large number of people by conducting non-consecutive, 2-day survey assessments (USDA, 2005b). Individual participants were asked to recall their dietary intake for the previous day on two separate occasions by an interview staff that adhered to a detailed survey protocol. The ﬁrst and second visits by the interview staff were separated by a time period of 3–10 days. The CSFII estimates are commonly combined with data regarding detectable pesticide residues per food type, such as the Pesticide Data Program (PDP) (USDA, 2005b) in order to approximate dietary pesticide exposure levels in the population (USEPA, 2000). Because patterns of dietary consumption, in contrast to snapshots of dietary intake, are especially important for estimating pesticide exposure levels among children, it is necessary to closely examine the methodology used to assess dietary intake. Therefore, the primary objective of this study was to apply the temporal sampling characteristics used in the cross-sectional approach to longitudinal food consumption data collected in the Children Pesticide Exposure Study in Washington (CPES-WA) (Lu et al., 2006) via a bootstrap sampling technique in order to investigate potential differences in the overall estimates of dietary intake. Moreover, we examined the variability over time as it relates to the precision of the intake estimate and explored the contribution of seasonality toward the dietary patterns of children in this study. 2. Methods 2.1. Study designs The longitudinal study, CPES-WA, was reported previously (Lu et al., 2006). In brief, children aged 3–11 were recruited from a school district in Washington State by means of an informational ﬂyer. The 23 children participating in this study were asked to maintain a food diary over a 7 to 15-day period in each of the four seasons over a time frame of one year. To maintain compatibility with the CSFII survey, food items consumed by children in the CPES-WA study were coded in accordance with the
CSFII protocol and grouped into the categories of fruits, grains, vegetables and juices. Because estimates of serving sizes were unreliable for the CPES-WA study and were obtained through recall in the crosssectional study, the frequency of reported intake was used as the outcome variable. Frequency of reported intake corresponds to the number of times an item of a particular food category was consumed per day per subject. CSFII data was extracted for children ranging in age from 3 to 11 residing in the US West region in order to match the demographic background of the CPES-WA participants. Within the CSFII cohort, signiﬁcant differences in food consumption frequencies were observed between income levels, and thus, the extracted CSFII data was further limited to children living in households with incomes of $80,183 or more, the high-income bracket deﬁned by the US Department of Health and Human Services in 1996 (500% of the poverty threshold for a family of four); which approximates the household income in the area where the CPES-WA study was conducted (Census, 2000). Children whose ﬁrst and second interview dates were not separated by 3–10 days, as stated in the CSFII study protocol, were also not included in this analysis (approximately 23% of total CSFII participants in this sub-group). In total, dietary intake information from 150 CSFII children was extracted for use in comparison to the CPES-WA cohort.
2.2. Statistical analyses The R statistical programming language (RDCT, 2005) was utilized to assess differences in distribution of time-averaged dietary intakes between the two study sampling approaches as well as to acquire information regarding changes in within and between-subject variability. To apply the cross-sectional sampling technique, utilized by CSFII, to the CPES-WA data set, a bootstrap technique (Efron and Tibshirani, 1993; Lee et al., 2004) was used in order to compare a random selection of two nonconsecutive days per subject to the longitudinal estimate pertaining to that subject. A particular day within each sampling season was randomly selected for each subject in the longitudinal dataset, and a second day was then randomly selected in the same season, differing from the ﬁrst sampled day by 3–10 days in either direction. The average consumption frequency over the two sampled days was calculated for each food group and each individual. This procedure was replicated 5000 times. Descriptive statistics, such as means, medians, and 95th percentiles of the bootstrap population distributions, were then calculated and compared to the longitudinal population parameters. A comparison of population means was performed using a one-way ANOVA and comparison of population distribution for each bootstrap replicate in reference to the longitudinal population distribution was assessed using the Kolmogorov–Smirnov test and F-test. The F-test assumes a normal distribution of the populations being compared and so, estimates were log transformed to better meet this assumption. Between-subject variability over time for the 15 consecutive days sampling season, as well as within-subject variability over consecutive days (i.e., 2-, 3-day, etc.) were determined and evaluated for signiﬁcant differences in standard error based on a paired T-test. Signiﬁcance was determined as a Po0.05.
3. Results The data generated by the bootstrapping approach allowed us to investigate potential differences in distribution of the food intake variables for the cross-sectional sampling scheme compared to the longitudinal distribution. This comparison suggests that the populationaveraged frequency of consumption does not differ between the longitudinal and bootstrap data. The mean frequency of intake for fruit, juice and vegetables was about once per day, while grains were consumed with a mean frequency of 3 times per day. However, the tails of the food consumption distributions appeared to differ
ARTICLE IN PRESS M.L. Givens et al. / Environmental Research 103 (2007) 325–330
between the longitudinal estimates and the bootstrap replicates (Fig. 1). To investigate this potential difference, we compared the population estimate for each bootstrap replicate to the longitudinal study population estimate using an F-test and the Kolmogorov–Smirnov test. The Ftest represents a more stringent or sensitive statistical test for comparison of population variances, while the Kolmogorov–Smirnov test is a more conservative test for comparison of overall population distributions. With these approaches we observed signiﬁcant differences in the two population estimates for the four categorized food variables. Table 1 indicates the percentage of signiﬁcant differences in population distributions comparing the
Average Total Intake Frequency
Fig. 1. Boxplots of the frequency of time-averaged total food intake for the longitudinal cohort in the CPES-WA study and representative replicates of the bootstrap cohorts. The boxplots denote the lower quartile, median value and upper quartile as well as the minimum and maximum values for average total food intake frequency. An ‘‘L’’ represents the distribution for the longitudinal cohort while each ‘‘B’’ designates the distribution for any one bootstrap replicate.
bootstrap replicates to the longitudinal estimate for each food category when employing the two statistical tests. Moreover, Table 1 provides the 50th and 95th percentile ranges observed from the bootstrap replicates as well as the corresponding estimates for the longitudinal data. By plotting the distribution of the two-day averages for the frequency of total food intake obtained from the bootstrap application (represented with a dashed line) as it compared to the longitudinal mean per subject (represented as an ‘x’), we examined this data from a different perspective. Fig. 2 shows that for any one subject, a wide range of values could have been calculated for the time-averaged total frequency of dietary intake based on a 2-day sampling scheme. Given the fact that an estimation of dietary intake frequency inherently has more variability when it is based on a cross-sectional, 2-day sampling scheme, as shown in Figs. 1 and 2, we explored the change in variability over time as the sampling period increases for individual subjects. Predictably, as more observations are added, the between-subject variation decreases, though differentially over time (Fig. 3), as well as the within-subject variation (Fig. 4). Variability over time within subjects for data collected consecutively for up to 5 days per subject was signiﬁcantly higher as compared to a 15 consecutive day sampling period (paired T-test, Po0.05). Conversely, estimates of variability obtained from 6 or more consecutive days of sampling did not statistically differ from the variability for 14 days of consecutive sampling. Because seasonality has been proposed to inﬂuence the consumption of food types (Shahar et al., 1999, 2001; Fowke et al., 2004), we investigated the plausibility of such a contribution in the longitudinal study population. When we compared the average frequency of intake for the food categories according to four seasons using a one-way ANOVA, we observed signiﬁcant differences in the mean intake frequency of fruit and the mean intake frequency of grain across seasons, but not for vegetables or juice in the longitudinal dataset (Table 2). Fig. 5 shows a graphical representation of the change in the average frequency of fruit and grain intake across seasons for the longitudinal study cohort. We observed a near mirror image in the trends for the average frequency of fruit intake vs. grain intake across seasons. These data suggest that seasonal differences may be important in the consideration of overall intake patterns.
Table 1 Comparison of population distributions for the longitudinal vs. cross-sectional sampling study design Food category
K–S test % of replicates
F-test % of replicates
Longitudinal 50th percentile
Bootstrap range of values 50th percentile
Longitudinal 95th percentile
Bootstrap range of values 95th percentile
Fruit Grain Juice Vegetables
2 2 2 21
0.18 94 2 48
1.00 3.64 0.50 0.71
0.50–2.00 2.50–4.50 0–1.50 0.50–1.50
3.52 4.46 2.61 1.58
1.45–5.35 4.00–6.90 1.00–4.45 1.45–4.00
Percentage of replicates denotes the number of bootstrap population distributions that signiﬁcantly differed from the longitudinal population distribution (Po0.05). Additional descriptive statistics denote values for reported frequency of food consumption.
ARTICLE IN PRESS M.L. Givens et al. / Environmental Research 103 (2007) 325–330
6 Standard Error
Average Total Intake Frequency
Fig. 2. Range of time-average total food intake frequency of bootstrap replicates vs. the longitudinal estimate for individual subjects. The dashed lines represent the range of estimates obtained using the bootstrap sampling technique and the ‘‘X’’ denotes the estimate obtained for the subject utilizing the longitudinal sampling data. On the X-axis, subjects are not arranged in numerical order.
7 8 Days
10 11 12 13 14
Fig. 4. Variability of total food intake frequency within subjects. All possible combinations of consecutive days were utilized to calculate within-subject variability. A signiﬁcant difference in variability was noted between the day 1 estimate and day 5 or greater day estimates (paired Ttest, Po0.05). No signiﬁcant difference was observed between the day 6 or greater estimates and the day 14 estimate.
Table 2 Consumption of fruit, grain, juice, and vegetable food types across seasons
Average intake frequency
7 8 Days
1.0970.20 1.2970.24 1.5970.22 1.2170.19
3.6670.18 3.5970.23 3.1670.19 3.5670.17
0.8270.16 0.7870.10 0.9770.15 0.8870.16
0.9270.12 0.8870.13 1.0470.13 1.1170.12
10 11 12 13 14
Fig. 3. Variability of total food intake frequency between subjects over time. The X-axis represents consecutive days of follow-up and the Y-axis denotes the standard error for the cohort over time.
Winter Spring Summer Fall
The average and standard error for the estimate of the frequency of intake for each food category are indicated. Signiﬁcant differences were determined by one-way ANOVA. **Po0.001, *Po0.05.
Exposure to pesticides among children in particular has recently been recognized as a signiﬁcant public health issue (Eskenazi et al., 1999). The risk among children is disproportionate to that of adults exposed to pesticides. Children need more protection due to their heightened level of exposure and the biological implications of such exposures, which warrants increased focus and additional study for this population. Paramount to the understanding of pesticide exposure among children is the need for
accurate estimates of dietary consumption patterns since a recent study has demonstrated that children are predominantly exposed to OP pesticides through dietary intake (Lu et al., 2006). Thus, we propose that the study design used to obtain an approximation of dietary intake can substantially change the estimated pesticide exposure levels. Given the nature of the exposure, the vulnerability of the population, and the observation that children are predominantly
ARTICLE IN PRESS M.L. Givens et al. / Environmental Research 103 (2007) 325–330
Average Frequency of Intake
3 Grain Fruit
Fig. 5. Mean food intake frequency for fruits and grains across four seasons for children participating in the CPES-WA longitudinal study. Circles and squares signify the mean frequency of intake and bars indicate the standard error for grains and fruits, respectively. The mean frequency of intake for fruits or grains signiﬁcantly differed by seasons using a oneway ANOVA.
exposed to certain pesticides such as OP pesticides through dietary intake, it is imperative to choose an appropriate temporal sampling approach to attain an accurate exposure assessment. An evaluation of the current methodology used to gather dietary intake information among children indicates that a cross-sectional study design may overestimate variability. We found that although there was no difference in the population-averaged frequency of consumption, whether assessed in a longitudinal or cross-sectional manner, there were signiﬁcant differences in the overall distribution of these estimates (Table 1). Indeed, such differences could signiﬁcantly change conclusions regarding regulatory standards if, for example, the 95th percentile cut point was used to calculate overall pesticide exposure levels (USEPA, 1992). The variability in observed dietary intake appears to be strongly inﬂuenced by the time period for follow-up. Because variability in any individual’s time-averaged intake estimate should decrease in proportion to the square root of the number of days of observation, precision should increase with longer follow-up over exchangeable days, but with diminshing marginal gains. Our results suggest that following an individual for at least 6 consecutive days may provide comparable estimates of dietary intake to those obtained from a 15 consecutive day follow-up period (Fig. 4). This conclusion is supported by previous ﬁndings (Buck et al., 1995; Lee et al., 2004), which suggest that longitudinal study designs may provide more accurate parameter estimates by substantially lowering the variability in estimated time-averaged dietary intake for any particular individual.
Results of dietary consumption from the CPES-WA study imply that the mean frequency of intake for fruits and grains signiﬁcantly varies by season. Thus, a snapshot of an individual’s dietary intake in one season, as would be obtained in a cross-sectional study design, may not reﬂect a year-round consumption pattern. The assumption that the CSFII survey makes when employing the cross-sectional study design is that one individual of a particular age, sex, income bracket, and regional residence is exchangeable for another individual of the same description across seasons. However, given the variability in a 2-day assessment, it may not be possible to obtain an accurate depiction of a child’s dietary habits, nor plausible to observe seasonal differences using a cross-sectional approach. The increased consumption of fresh fruits in the summer and the subsequent decrease in the fall and winter seasons, as observed in the longitudinal data, corresponds well with the availability of fresh fruits in these seasons. It should be noted that our longitudinal cohort consisted of children aged 3–11 from families in the high-income bracket as deﬁned by the US Department of Health and Human Services, residing in Washington State and therefore results are not necessarily relevant for all ages, demographics or residential locations. Because the total amount of food per body weight consumed by children is the same across the four seasons in our cohort, the consumption patterns for grain-based foods, such as cereal, mufﬁns, or pancakes, actually complements the consumption pattern for fruits in this cohort. Due to detectable pesticide residues measured in fruits and grain-based food items (USDA, 2005b), periodic differences in dietary consumption imply that children’s exposure to pesticides via dietary intake may be inﬂuenced by a seasonal effect. Although the pools from which subjects are sampled in the CPES-WA study and the nationwide CSFII survey differ, the temporal sampling technique or bootstrap approach applied to the longitudinal study cohort mimics that of the CSFII sampling protocol. Therefore, each boot replicate can be viewed as a simulated CSFII survey assessment of the longitudinal study cohort. Indeed, we noted that the estimates obtained from the bootstrap technique resemble the distribution obtained from the CSFII sub-cohort, matched by age, residential location, and income. Thus, it is plausible to suggest that the 2-day sampling scheme inherently increases the variability of national dietary intake estimates as compared to a longitudinal approach, as was observed in our simulated results (Figs. 1–4, Table 1). Despite concerns regarding variability, a cross-sectional study design with a sufﬁcient sample size can approximate the population mean and median estimates generated from the longitudinal approach. Therefore, although the tails of the distributions for food intake by children obtained from the CSFII survey may not offer consistent estimates for calculating exposure assessments, this survey can be used for approximating the population-averaged food consumption.
ARTICLE IN PRESS 330
M.L. Givens et al. / Environmental Research 103 (2007) 325–330
In conclusion, it is apparent that differences exist in the ultimate parameter estimates dependent upon the survey sampling methodology applied for collecting dietary consumption data from children. The variability that arises as a result of a 2-day, cross-sectional approach will signiﬁcantly alter the tails of consumption distributions, which are commonly used for cumulative risk assessment or regulatory purposes to calculate dietary pesticide exposures. Given the vulnerability to pesticide toxicity in children and the nature of the exposure via dietary intake, it is imperative that variability be minimized in order to improve the accuracy of estimating food consumption and the subsequent pesticide exposures via dietary intake among children. Data obtained from a longitudinal sampling approach in which dietary consumption information would be collected for at least 6 days in each of the 4 seasons would not only decrease variability, but also identify other factors, such as seasonality, that may explain additional variability in parameter estimates. Acknowledgments This work was supported by the US Environmental Protection Agency (EPA), Science to Achieve Results (STAR) program (RD-831539). Its contents are solely the responsibility of the authors and do not necessarily represent the ofﬁcial view of US EPA. M. Givens contributed to compilation of the data, developed and conducted the statistical approach, and wrote the manuscript. C. Lu collected the longitudinal data and developed the methodological comparison concept. S. Bartell developed the statistical approach and consulted on R coding. M. Pearson compiled the data set and consulted on statistical approaches. In addition, we thank C. Holbrook, K. Conlon, M. Mehta, and H.-S. Hsiao for constructive commentary, as well as T.K. Sato for critical reading of the manuscript. Dietary consumption information was obtained from children participating in the Children Pesticide Exposure Study (CPES-WA). Written consent or oral assent was obtained from parents and older children, or from younger children, respectively. The University of Washington Human Subject Committee approved the use of human subjects in this study. References Buck, R.J., Hammerstrom, K.A., Ryan, P.B., 1995. Estimating long-term exposures from short-term measurements. J. Expo. Anal. Environ. Epidemiol. 5, 359–373.
Census US, 2000. US Census Bureau. Fact Sheet: Census 2000 Demographic Proﬁle HIghlights. Curl, C.L., Fenske, R.A., Elgethun, K., 2003. Organophosphorus pesticide exposure of urban and suburban preschool children with organic and conventional diets. Environ. Health Perspect. 111, 377–382. Daniels, J.L., Olshan, A.F., Savitz, D.A., 1997. Pesticides and childhood cancers. Environ. Health Perspect. 105, 1068–1077. Dharmani, C., Jaga, K., 2005. Epidemiology of acute organophosphate poisoning in hospital emergency room patients. Rev. Environ. Health 20, 215–232. Efron, B., Tibshirani, R.J., 1993. An Introduction to the Bootstrap. Chapman & Hall/CRC, London, Boca Raton, FL. Eskenazi, B., Bradman, A., Castorina, R., 1999. Exposures of children to organophosphate pesticides and their potential adverse health effects. Environ. Health Perspect. 107 (Suppl. 3), 409–419. Fenske, R.A., Curl, C.L., Kissel, J.C., 2003. The effect of the 14-day agricultural restricted entry interval on azinphosmethyl exposures in a group of apple thinners in Washington state. Regul. Toxicol. Pharmacol. 38, 91–97. Fowke, J.H., Schlundt, D., Gong, Y., Jin, F., Shu, X.O., Wen, W., Liu, D.K., Gao, Y.T., Zheng, W., 2004. Impact of season of food frequency questionnaire administration on dietary reporting. Ann. Epidemiol. 14, 778–785. Lee, K., Bartell, S.M., Paek, D., 2004. Interpersonal and daily variability of personal exposures to nitrogen dioxide and sulfur dioxide. J. Expo. Anal. Environ. Epidemiol. 14, 137–143. Lu, C., Toepel, K., Irish, R., Fenske, R.A., Barr, D.B., Bravo, R., 2006. Organic diets signiﬁcantly lower children’s dietary exposure to organophosphorus pesticides. Environ. Health Perspect. 114, 260–263. NRC, 1993. Pesticides in the Diets of Infants and Children. National Academy Press, Washington, DC. RDCT, 2005. R: a language and environment for statistical computing. In: ISBN 3-900051-07-0 Edition. R Foundation for Statistical Computing: R Development Core Team, Vienna, Austria. Repetto, R., Baliga, S.S., 1996. Pesticides and the immune system: the public health risks. Executive summary. Cent. Eur. J. Public Health 4, 263–265. Shahar, D.R., Froom, P., Harari, G., Yerushalmi, N., Lubin, F., KristalBoneh, E., 1999. Changes in dietary intake account for seasonal changes in cardiovascular disease risk factors. Eur. J. Clin. Nutr. 53, 395–400. Shahar, D.R., Yerushalmi, N., Lubin, F., Froom, P., Shahar, A., KristalBoneh, E., 2001. Seasonal variations in dietary intake affect the consistency of dietary assessment. Eur. J. Epidemiol. 17, 129–133. USDA, 2005a. US Department of Agriculture Pesticide Data Program Annual Summaries (1994–2000). /http://wwwamsusdagov/science/ pdpS. USDA, 2005b. Food Surveys Research Group. Beltsville Human Nutrition Research Center. United States Department of Agriculture. USEPA, 1992. Guidelines for Exposure Assessment. US Environmental Protection Agency, Risk Assessment Forum, Washington, DC, 600Z92/001. USEPA, 2000. Choosing a percentile of acute dietary exposure as a threshold of regulatory concern. Ofﬁce of Pesticide Programs, United States Environmental Protection Agency. Zahm, S.H., Ward, M.H., 1998. Pesticides and childhood cancer. Environ. Health Perspect. 106 (Suppl. 3), 893–908.