Validation and relative validation studies of dietary reporting methods are conducted to evaluate the extent to which those methods elicit accurate reports; this involves comparing reports obtained by such methods as 24-hour recalls and food-frequency questionnaires (FFQs) with reference information obtained by such methods as direct observation, duplicate portion collection or food records. The reported and reference information each consist of a set of food items and their respective amounts. (Table 1 shows terms used in this article and their definitions.)
Table 1 Terminology
Assessment of reporting accuracy for a group of participants often begins with converting both reported and reference information to energy and nutrients. After cumulating, for each participant, the energy and nutrients within each set, analyses involve several descriptive measures and inferential tests (see e.g. references Reference Masson, McNeill, Tomany, Simpson, Pearce and Wei1–Reference Todd and Kretsch14). First, paired t-tests may be used to compare the mean of the participant-specific differences between reported and reference energy and nutrients with 0. Second, the correlation, over participants, between reported and reference energy and nutrients may be calculated, with correlations near 1 presumed to indicate higher reporting accuracy. Third, the ratio of reported to reference energy and nutrients, multiplied by 100, may be calculated. We call this the report rate; others have labelled it ‘percentage recalled’ and ‘percentage over (or under) reported’ (see e.g. references Reference Bransby, Daubney and King3, Reference Conway, Ingwersen, Vinyard and Moshfegh5 and Reference Conway, Ingwersen and Moshfegh6) Values close to 100%, found when reported and reference energy and nutrients are approximately equal, are presumed to represent high reporting accuracyReference Reynolds, Johnson and Silverstein12–Reference Todd and Kretsch14. We call these analyses, taken together, the ‘conventional approaches’ to evaluating reporting accuracy.
These conventional approaches are generally indifferent to whether food items and amounts are reported correctly – reported information is converted to energy and nutrients regardless of whether items were actually eaten. In this article, we demonstrate that conventional analyses of data from dietary reporting validation studies may mask the complexity of dietary reporting error and overestimate reporting accuracy.
Consider, for example, mean differences between reported and reference information. For any dietary measure, the mean difference would be 0 in a validation study in which all participants reported items and amounts with perfect accuracy. However, the mean difference could also be 0 if, on average, participants reported items and amounts that they did not eat that offset their failures to report items and amounts that they did eat. Suppose that an individual was observed eating a serving of hash browns, and subsequently reported eating a serving of apple sauce. Converted to energy and macronutrients, the individual was observed eating 100 kcal, 1 g of protein, 12 g of carbohydrate and 5 g of fat; and reported eating 96 kcal, 0.2 g of protein, 25.4 g of carbohydrate and 0.2 g of fat. (Composition information for a 57 g serving of hash browns is from the product label, and for a 120 ml serving of sweetened apple sauce is from the Nutrition Data System for Research [NDSR, version 4.03; Nutrition Coordinating Center, University of Minnesota] database.) The report rates for energy (96%), protein (20%), carbohydrate (212%) and fat (4%) would be interpreted as essentially perfect reporting of energy, underreporting of protein and fat, and overreporting of carbohydrate. However, all reported energy and macronutrients were from a reported item that was, in fact, not eaten by that individual. The conventional approaches, by their indifference to whether reported items and amounts were actually eaten, may misrepresent dietary reporting accuracy.
We introduce some terminology to facilitate our discussion of this problem and of our recommendations concerning how to address it. We then present an illustration using data from a validation study of dietary recall, and so frame the discussion in terms of dietary recall. However, the concerns also pertain to relative validation of FFQs.
In dietary recall, the reported information is a set of item–amount pairs (e.g. [apple, 1 medium]; [whole milk, 1 serving]) reported by a participant. The reference information is also a set of item–amount pairs.
Reporting accuracy may be evaluated by assessing the congruence between the reported and reference sets of information, which involves partitioning the sets as illustrated in Fig. 1 and discussed here.
Fig. 1 Dietary reporting accuracy in a validation (or relative validation) study should be assessed by evaluating the congruence between reported and reference information. In this approach, reported and reference items are classified as intrusions (reported items not in the reference set), matches (reported items in the reference set) and omissions (reference set items not reported). Amounts – of servings or of energy and nutrients – are classified as overreported (reported but not in the reference information), corresponding (reported and in the reference information) and unreported (in the reference information, but not reported)
For any participant, reported items are partitioned into matches and intrusions, and reference items are partitioned into (the same) matches and omissionsReference Smith15, Reference Smith, Jobe and Mingay16. A match is an item reported eaten that is in the reference set. An intrusion is an item reported eaten that is not in the reference set. An omission is an item in the reference set that is not reported eaten.
Amounts – whether of servings or of such quantities as kilocalories and grams of protein – are classified as corresponding, overreported or unreported. For a match, the reported amount equals, exceeds or falls short of the reference amount. If the reported amount equals the reference amount, the reported amount is a corresponding amount. If the reported amount exceeds the reference amount, the amount by which it does so is an overreported amount; if the reported amount falls short of the reference amount, the amount by which it falls short is an unreported amount. In each case, the overlap between the reported and reference amounts is a corresponding amount. (For example, if an individual was observed consuming a 1/4 serving of milk, and reported consuming a whole serving, the reported serving consists of a corresponding 1/4 serving and an overreported 3/4 serving. If an individual was observed consuming 1 serving of milk and reported consuming 1/4 serving, the reported 1/4 serving is corresponding, and 3/4 serving is unreported.) For an omission, the entire amount is unreported. For an intrusion, the entire amount is overreported.
From this, it follows that:
and
in which each term is an arithmetic quantity and + represents addition. Letting rep, ref, c and o represent reported, reference, corresponding and overreported amounts, respectively, we can abbreviate the former expression to rep = c+o.
The report rate of energy or of a nutrient, a conventional measure of reporting accuracy, is the reported amount divided by the reference amount, multiplied by 100. A report rate has a lower bound of 0%, which would indicate that nothing was reported, and no upper bound, because there is no limit on what an individual might report. Algebraic expression of the report rate, as
![( rep / ref )\times 100 = [( c + o )/ ref ]\times 100 = [( c / ref )\times 100] + [( o / ref )\times 100],](https://static.cambridge.org/binary/version/id/urn:cambridge.org:id:binary:20151022095957546-0642:S136898000768714X_eqnU3.gif?pub-status=live){kind=small}
shows that it consists of two parts. We call [(c/ref) × 100] the correspondence rate: it is the percentage of the reference amount to which the reported amount corresponds. It is a genuine measure of accuracy with a value between 0%, which would indicate that no reference item had been reported, and 100%, which would indicate that all reference items and their amounts had been reported correctly. We call [(o/ref) × 100] the inflation ratio: it is a non-negative augmentation to correctly reported information based on inaccurate reporting; it has no upper bound, because there is no limit on what can be reported.
Although the report rate is treated conventionally as a measure of accuracy, it includes a component based on reporting error: because the inflation ratio is non-negative, any overreporting – whether it is reporting intrusions or amounts of matches that exceed reference values – inflates the report rate.
Consider again the individual who was observed eating one serving of hash browns and reported eating one serving of apple sauce, and assume that this individual ate only hash browns and reported only apple sauce. This individual's report rates for energy, protein, carbohydrate and fat, are 96%, 20%, 212% and 4%, respectively, but the correspondence rate for each is 0%, because none of what was reported was from a match. The inflation ratios for energy and each macronutrient are equal to the respective report rates, because all of what was reported was from an intrusion.
In this article, we demonstrate that conventional analyses of data from dietary reporting validation studies that convert reported and reference information to energy and nutrients, and that do not distinguish between reported matches and intrusions or between corresponding and overreported amounts, may mask the complexity of dietary reporting error and overestimate reporting accuracy. We use data from a study conducted to determine whether children's dietary recall accuracy depends on interview modality (in-person or by telephone)Reference Baxter, Thompson, Litaker, Guinn, Frye and Baglio17.
Methods
Our objective was to compare conclusions that might be drawn from different analytical approaches to dietary validation data, so the particularities of the original study are of minimal interest. We described the sample and methodology elsewhereReference Baxter, Thompson, Litaker, Guinn, Frye and Baglio17, so summarise them here.
The Institutional Review Board of the Medical College of Georgia approved the procedures for recruiting participants and collecting data, and that of the University of South Carolina approved using the data for the analyses described in this article.
Participants
During the 2001–2002 school year, all fourth graders (n = 799) from 10 public schools in one school district were invited to participate. From the 451 children who provided written child assent and parental consent, children were selected randomly, subject to the constraint that there be roughly equal numbers in each of four race (black, white) × sex groups. Sixty-nine children (35 girls) were interviewed; their mean age at the time of their interviews was 10.15 years (standard deviation = 0.55 years).
Procedures
To compile reference information, one of three dietitians observed each child eating school breakfast and school lunch on a school day. Following procedures used in previous studiesReference Baxter, Thompson, Litaker, Frye and Guinn18, Reference Baxter, Thompson, Smith, Litaker, Yin and Frye19, observers recorded items eaten and their amounts in servings. Only children who obtained breakfast and lunch at school were observed, because it is difficult to identify contents of meals brought from homeReference Simons-Morton, Forthofer, Huang, Baranowski, Reed and Fleishman20. Children were observed for their entire meal periods so that any food trading could be notedReference Crawford, Obarzanek, Morrison and Sabry7, Reference Simons-Morton and Baranowski21, Reference Baxter, Thompson and Davis22. Although children knew in general that they were being observed, they did not know who, if anyone, would be interviewed or whether any interview would be in person or by telephone. Weekly assessments indicated that interobserver reliability was satisfactoryReference Baxter, Thompson, Litaker, Guinn, Frye and Baglio17, Reference Baglio, Baxter, Guinn, Thompson, Shaffer and Frye23.
Within each race × sex group, the children were divided randomly between the in-person and telephone interview modalities. One of two dietitians interviewed each child after 6.30 p.m. on the day that the child had been observed; the interviewer had not observed either of the child's meals. During the interview, the child was to describe and quantify everything consumed that day; this was that child's reported information. Interviewers followed a protocol modelled on the NDSR computerised interview, but wrote information on forms. Interviews were audiorecorded. Interview quality was monitored by daily assessment of a randomly selected interview on a structured list of criteria, including agreement among the interview form, the audiotape and the typed transcriptReference Shaffer, Baxter, Thompson, Baglio, Guinn and Frye24; these assessments indicated that interviewers adhered to protocolReference Baxter, Thompson, Litaker, Guinn, Frye and Baglio17.
Construction of analytic variables
Reference information was available only for school breakfast and school lunch, so analyses were restricted to reports of those meals.
We prepared variables for two analytic approaches; each involved converting reported and reference information to energy and macronutrients. (1) For conventional analyses, we ignored whether reported items were matches or intrusions (and therefore whether reported amounts were corresponding or overreported). (2) For analyses sensitive to reporting errors, we classified reported items as matches or intrusions, and reported amounts as corresponding or overreported.
We published results concerning reporting accuracy at the food-item level elsewhereReference Baxter, Thompson, Litaker, Guinn, Frye and Baglio17. These showed that none of omission rate (percentage of observed items not reported; overall mean = 33%), intrusion rate (percentage of reported items not observed; overall mean = 17%) and total inaccuracy (sum of itemwise absolute difference between observed and reported numbers of servings; overall mean = 4.4 servings) depended on whether interviews were conducted in person or by telephoneReference Baxter, Thompson, Litaker, Guinn, Frye and Baglio17.
As in our previous studies, reported meals were treated as reports about school meals only if the child reported eating that meal at school; referred to breakfast as ‘school breakfast’ or ‘breakfast’ and to lunch as ‘school lunch’ or ‘lunch’; and reported the mealtime to within an hour of the observed mealtimeReference Baxter, Thompson, Litaker, Guinn, Frye and Baglio17–Reference Baxter, Thompson, Smith, Litaker, Yin and Frye19.
Numeric values of servings were assigned, as follows, to the qualitative labels used during observations and interviews: none (0.00), taste (0.10), little bit (0.25), half (0.50), most (0.75), all (1.00) or the actual number of servings if more than one was observed or reportedReference Baxter, Thompson, Litaker, Guinn, Frye and Baglio17–Reference Baxter, Thompson, Smith, Litaker, Yin and Frye19. For each reference item and each reported item, we obtained per serving information about energy and macronutrient content from the NDSR database or, for items not in that database, from product information or recipes provided by the school district's nutrition programme.
Conventional variables
Reference and reported amounts of energy and macronutrients were calculated from reference and reported information, respectively, and food composition information by multiplying the quantified servings of each item by the per-serving energy and macronutrient values. For each child, for energy and each macronutrient, these values were summed across the reference (i.e. observed) items for the two school meals, and also across the items reported for the two school meals. Table 2 illustrates, for one child, how we generated values of reported kilocalories and reference kilocalories – the variables required for conventional analyses of energy reporting accuracy. For each child, for energy and each macronutrient, we calculated the report rate from values of these variables.
Table 2 Classifications and computations used to assess accuracy of reported energy compared with reference energy for one child*
*Reference information was obtained by observation. This child was observed eating chocolate milk, sausage and grits at school breakfast, and honey mustard, chocolate milk, rice pilaf and egg roll at school lunch. This child reported eating chocolate milk, peaches and grits at school breakfast, and chocolate milk, rice pilaf and broccoli at school lunch.
† Every item that was reported eaten or observed eaten or both is listed.
‡ From the Nutrition Data System for Research database (NDSR, version 4.03; Nutrition Coordinating Center, University of Minnesota) or from product information or recipes obtained from the school district's nutrition programme.
§ Values are products of serving and kcal per serving information. Reported kcal are either corresponding kcal or are overreported, so that in each row, corresponding kcal and overreported kcal sum to reported kcal. Reference kcal are either corresponding or unreported, so that in each row, corresponding kcal and unreported kcal sum to reference kcal.
¶ Overreported kcal from intrusions = kcal from intrusions (i.e. food items reported eaten but not in reference set).
‖ Overreported kcal from matches = for matches for which reported kcal > reference kcal, absolute differences between reported kcal and reference kcal.
**Corresponding kcal from matches = for matches, overlap between reported kcal and reference kcal. (It is the minimum of reported kcal and reference kcal. Corresponding amounts are necessarily from matches.)
†† Unreported kcal from matches = for matches for which reported kcal < reference kcal, absolute differences between reported kcal and reference kcal.
‡‡ Unreported kcal from omissions = kcal from omissions (i.e. food items in reference set but not reported eaten).
§§ This item was observed on this child's tray but none was eaten. The child mentioned this item during the interview but reported that none of the item was eaten. This item was not included in calculating reporting accuracy.
¶¶ This item was observed on this child's tray but none was eaten and it was not reported in the interview. This item was not included in calculating reporting accuracy.
Variables sensitive to reporting errors
For each child, following classification of each reported and reference item as a match, an intrusion or an omission, the constituent energy and macronutrients of each item were classified as corresponding, overreported or unreported.
An item reported by a child was treated as matching a reference item unless the reported item clearly did not describe an item that the child was observed eating; this might overestimate accuracy. For example, any reported white milk (e.g. skimmed, whole) was matched to any observed white milk, and any reported pizza (e.g. cheese, pepperoni) was matched to any observed pizza. However, reported fruit juices (e.g. apple, orange), vegetables (e.g. green beans, broccoli) and milk flavours (e.g. chocolate, strawberry) that differed from what was observed were not classified as matchesReference Baxter, Thompson, Litaker, Guinn, Frye and Baglio17–Reference Baxter, Thompson, Smith, Litaker, Yin and Frye19.
Each corresponding, overreported and unreported number of servings was multiplied by the appropriate per-serving values of kilocalories and grams of macronutrients to obtain corresponding, overreported and unreported amounts of energy and macronutrients. For each child, summing the values for each category across items yielded total kilocalories and total grams of each macronutrient that were corresponding, overreported and unreported. Table 2 illustrates, for energy for one child, how we generated these variables. For each child, for energy and each macronutrient, we used appropriate values of these variables to calculate the correspondence rate and the inflation ratio.
Analyses
Analyses were conducted with SAS (version 8.2; SAS Institute).
Effect of interview modality
Whether interviews were in person or by telephone was irrelevant to the concerns of this article but, before conducting analyses for this article, we compared the interview-modality groups on correspondence rates and report rates for energy and each macronutrient (data not shown). Using Wilcoxon rank-sum tests, we found no significant effect of interview modality on either measure for energy or any macronutrient (every P>0.14), so, for subsequent analyses, we combined data from the two groups.
Conventional variables
We used t-tests to compare mean differences between reported and reference values of energy and of each macronutrient with 0. We calculated Pearson correlations, over subjects, between reported and reference energy and macronutrient values. (Neither the t-tests nor the tests of correlation coefficients were independent, because the energy and macronutrient variables were not independent – all variables were calculated from single sets of food items. However, treating such variables separately appears to be customary practice.) For each macronutrient, we also calculated correlations between energy-normalised reference values (grams observed/kilocalories observed) and energy-normalised reported values (grams reported/kilocalories reported).
Variables sensitive to reporting errors
Because the distributions of inflation ratios were right skewed, we used sign tests to compare inflation ratios for energy and for each macronutrient with 0.
Results
Conventional variables
The first two data columns of Table 3 show descriptive statistics for reported and reference amounts of energy and each macronutrient. For energy and every macronutrient, the mean reported amount was significantly less than the mean reference amount (for energy, protein, carbohydrate and fat, paired t(68)s = 4.57, 2.37, 4.35 and 5.00, respectively; every P < 0.021). However, for energy and every macronutrient, the correlation, over children, between reported and reference amounts was statistically significant (for energy, protein, carbohydrate and fat, r = 0.56, 0.63, 0.52 and 0.61, respectively; every P < 0.001). Correlations between energy-normalised reported and reference values of protein, carbohydrate and fat were similar in magnitude (every r>0.48). The first data column of Table 4 shows that for energy and the macronutrients, median report rates ranged from 76% to 95%.
Table 3 Descriptive statistics for amounts of energy and macronutrients according to reported, reference, and five categories of amounts given as mean (standard deviation) (n=69)
*Reference information was obtained by observation. See Fig. 1 for an illustration of the five categories of amounts.
Table 4 Descriptive statistics for report rates, correspondence rates and inflation ratios for energy and macronutrients, given as median % (minimum %; first quartile %; third quartile %; maximum %) (n=69)
*For an individual, report rate = (Reported amount/Reference amount) × 100. A report rate has a lower bound of 0% (which indicates nothing was reported), but no upper bound (because there is no limit on what an individual can report).
† For an individual, correspondence rate = (Corresponding amount/Reference amount) × 100; this is the proportion of the reference amount to which the reported amount corresponds, and is a genuine measure of accuracy. A correspondence rate is between 0% (which indicates that nothing reported eaten had been observed eaten) and 100% (which indicates that all reported items were observed, with the number of servings being reported correctly). Corresponding amounts are necessarily from matches.
‡ For an individual, inflation ratio = (Overreported amount/Reference amount) × 100. Note that the numerator quantity is not a subset of the denominator quantity. The inflation ratio is non-negative and can exceed 100%. Overreported amounts may come from matches or from intrusions.
Variables sensitive to reporting errors
The last five data columns of Table 3 show descriptive statistics for the decomposition of reported and reference amounts of energy and each macronutrient into the five categories of amounts shown in Fig. 1. These values clarify that corresponding amounts from matches constituted only part of reported amounts, and that overreported amounts and unreported amounts were not in balance.
The last two data columns of Table 4 show descriptive statistics for correspondence rates and inflation ratios for energy and each macronutrient. Median correspondence rates were between 67% and 79%, and median inflation ratios were between 7% and 17%. Not surprisingly, median correspondence rates were lower than median report rates. Although minimum inflation ratios were 0% for energy and every macronutrient, all first quartiles were non-zero, indicating that many children reported intrusions, overreported amounts of matches, or both. Sign tests showed, for energy and for every macronutrient, that inflation ratios significantly exceeded 0 (every P < 0.0001).
Discussion
Conventional analyses provided a mixed picture of children's reports: although t-tests showed that children underreported their dietary intake, reported and reference information was significantly associated. Further, from median report rates for energy, protein and carbohydrate that were close to 100%, one might conclude that children, on average, reported these accurately. Although report rates of 100% have been interpreted as indicating perfect reporting accuracyReference Reynolds, Johnson and Silverstein12–Reference Todd and Kretsch14, given that report rates have no upper limit, and that their numerator quantities are not necessarily subsets of their denominator quantities, report rates have little logical connection to reporting accuracy.
Median energy and macronutrient correspondence rates for these children – from reported items that actually matched reference items – were lower, ranging from 67% to 79%. Reporting accuracy was worse than the conventional report rates suggested. The differences came from intrusions and overreported amounts of matches. These are quantified by inflation ratios, the median values of which ranged from 7% to 17%.
Overreported amounts – from intrusions and from matches – were not balanced by unreported amounts. Conventional analyses of validation study data that do not distinguish between reported matches and intrusions or between corresponding and overreported amounts provide an overly simplified and possibly misleading picture of dietary reporting accuracy.
Mixed findings have been yielded by validation studies that have obtained reference information by observation and used conventional approaches to assess the dietary-reporting accuracy for energy and macronutrients of 7- to 13-year-old children who reported without help from parentsReference Carter, Sharbaugh and Stapell4, Reference Crawford, Obarzanek, Morrison and Sabry7–Reference Lytle, Murray, Perry and Eldridge10, Reference Reynolds, Johnson and Silverstein12–Reference Todd and Kretsch14. Some studies have not found significant differences between reported and observed intake for energyReference Crawford, Obarzanek, Morrison and Sabry7, Reference Samuelson13Reference Todd and Kretsch14 (13-year-oldsReference Samuelson13; Chinese, HispanicReference Todd and Kretsch14), proteinReference Crawford, Obarzanek, Morrison and Sabry7, Reference Samuelson13Reference Todd and Kretsch14 (8-year-oldsReference Samuelson13; Chinese, Filipino, HispanicReference Todd and Kretsch14), carbohydrateReference Crawford, Obarzanek, Morrison and Sabry7Reference Todd and Kretsch14 (Chinese, HispanicReference Todd and Kretsch14) and fatReference Emmons and Hayes8,Reference Samuelson13Reference Todd and Kretsch14 (13-year-oldsReference Samuelson13; Chinese, HispanicReference Todd and Kretsch14), which could be interpreted as high reporting accuracy. Others have found reported intake to be significantly greater than observed intake for energyReference Lytle, Murray, Perry and Eldridge10, Reference Samuelson13Reference Todd and Kretsch14 (8-year-oldsReference Samuelson13; FilipinoReference Todd and Kretsch14), carbohydrate (FilipinoReference Todd and Kretsch14) and fatReference Samuelson13Reference Todd and Kretsch14 (8-year-oldsReference Samuelson13; FilipinoReference Todd and Kretsch14), which is conventionally interpreted as overreporting. Yet other studies have found reported intake to be significantly less than observed intake for energyReference Carter, Sharbaugh and Stapell4, Reference Reynolds, Johnson and Silverstein12,Reference Todd and Kretsch14 (CambodianReference Todd and Kretsch14), proteinReference Carter, Sharbaugh and Stapell4, Reference Samuelson13Reference Todd and Kretsch14 (13-year-oldsReference Samuelson13; CambodianReference Todd and Kretsch14) and carbohydrate and fatReference Reynolds, Johnson and Silverstein12Reference Todd and Kretsch14 (CambodianReference Todd and Kretsch14), which is conventionally interpreted as underreporting. One study found that although reported energy intake significantly exceeded reference intake, the percentages of energy from protein, carbohydrate, and fat in reported and observed items did not differ significantlyReference Lytle, Nichaman, Obarzanek, Glovsky, Montgomery and Nicklas9. Without distinguishing between matches and intrusions, and between corresponding and overreported amounts of matches, it is difficult, if not impossible, to assess the accuracy of children's reports in these studies. Regardless of whether energy and macronutrient values from reported information were greater than, less than, or not significantly different from values from reference information, reports could include intrusions and overreported amounts.
The correlation, over individuals, between energy and macronutrient values converted from reported and reference information is another conventional metric of reporting accuracyReference Crawford, Obarzanek, Morrison and Sabry7–Reference Lytle, Murray, Perry and Eldridge10, Reference Samuelson13, Reference Todd and Kretsch14. Correlations close to 1 are interpreted conventionally as indicating high reporting accuracy. In the data described in this article, correlations between reported and reference energy and macronutrients ranged from 0.52 to 0.63 (every P < 0.001). These correlations are similar to those found in many validation studies of dietary reports by 8- to 13-year-old children who reported without help from parents and in which reference information was obtained by observationReference Crawford, Obarzanek, Morrison and Sabry7–Reference Lytle, Murray, Perry and Eldridge10, Reference Samuelson13, Reference Todd and Kretsch14. However, correlations give no information about actual reporting accuracy: although, over a set of individuals, reported energy and macronutrients might covary with reference energy and macronutrients, respectively, there is no necessary connection, within individuals, between information obtained by aggregating over reports, on the one hand, and information obtained by aggregating over observations, on the other.
Epidemiological interest may often be in nutrients, not foods, so that failing to distinguish intrusions from matches, or overreported amounts from corresponding amounts, might not seem problematic. However, there is increasing interest in analysis of dietary patterns and their relationship to health outcomes (see e.g. references Reference Hu25–Reference Kant27) increasing appreciation that sources of nutrients are important and that nutrients interact (see e.g. references Reference Choi, Willett, Stampfer, Rimm and Hu28–Reference Baech, Hansen, Bukhave, Jensen, Sorensen and Kristensen31) and recognition that differential reporting accuracy for different foods on FFQs should be taken into account in applying measurement error correction methodsReference Rosner and Gore32. These scientific trends suggest that assessment of validity should be concerned with reporting accuracy of food items and their quantities.
In addition, as shown in Table 3, overreported amounts do not necessarily balance unreported amounts. The degree to which these amounts are out of balance likely depends on specific aspects of the interview. For example, we have found, with both adultsReference Smith, Jobe and Mingay16 and childrenReference Baxter, Smith, Guinn, Thompson, Litaker and Baglio33, that certain reporting instructions lead research participants to report more intrusions than do other instructions without affecting reports of matches. In conventional analyses, this would be manifested as higher report rates with the former instructions than with the latter, whereas analyses of variables sensitive to reporting errors would show that correspondence rates would be the same with the two instructions. How incorrect reports of amounts are distributed across the categories illustrated in Fig. 1 is an aspect of reporting complexity that is lost in conventional analyses of validation-study data, but that would be revealed by analyses of variables sensitive to reporting errors.
This investigation has certain limitations inherited from the design of the original study: these include that only fourth-grade children from one school district were studied; that observation was restricted to children who obtained their breakfast and lunch from the school food service; and that analysis was restricted to the school meal portions of children's reports.
Conclusions
Conventionally, reported information and reference information in dietary validation (and relative validation) studies are converted to energy and nutrients; t-tests are used to examine differences, and correlations are used to describe relationships, between reported and reference energy and nutrient values; and report rates are calculated. But these approaches do not take into account that reports of food items include intrusions as well as matches, that reported amounts of matches may be incorrect, and that reported amounts of intrusions are necessarily incorrect. Analyses that do not distinguish between matched and intruded food items, or between corresponding and overreported amounts, may provide a misleading picture of reporting accuracy.
It is often implicit – and sometimes explicit – in discussions of the results of validation studies that significant differences between reported and reference energy and nutrients indicate overreporting or underreporting of amounts, and that the absence of differences indicates excellent reporting accuracy. However, in the absence of knowledge of the correspondence between reported and reference information, it is impossible to estimate genuine reporting accuracy.
We suggest that conclusions about reporting accuracy from dietary validation (and relative validation) studies should not be drawn from energy and nutrient variables based on total reported and total reference information. Instead, we recommend analysis of energy and nutrient variables sensitive to reporting errors, constructed following classification of reported items as matches and intrusions, and of reported amounts as corresponding and overreported.
Acknowledgements
Sources of funding: This research was supported by grants 43-3AEM-2-80101 from the United States Department of Agriculture, Economic Research Service, Food Assistance and Nutrition Research Program, and R01 HL63189 from the National Heart, Lung, and Blood Institute of the National Institutes of Health. S.D.B. was Principal Investigator of both grants. A.F.S. was Principal Investigator of a subcontract of 43-3AEM-2-80101.
Conflict of interest declaration: No author has any conflict of interest.
Authorship responsibilities: Study concept–A.F.S., S.D.B.; acquisition of data, recruitment of schools and participants, data coding and entry, quality control–S.D.B.; statistical analysis and interpretation of data–A.F.S., S.D.B., J.W.H., M.D.N.; drafting and revising the manuscript–A.F.S., S.D.B., J.W.H., M.D.N.
Acknowledgements: Some of these results were presented at the Sixth International Conference on Dietary Assessment Methods (Copenhagen, 2006). We appreciate the cooperation of children, faculty and staff at the schools at which we collected data, and of the school district's nutrition programme and board of education. We thank Ram Chandran (Economic Research Service, USDA), Katherine Flegal (National Center for Health Statistics) and Caroline H Guinn (University of South Carolina) for their comments on earlier versions of this article, and J Whitney Keener (Cleveland State University) for her assistance in preparing the manuscript. We thank each named individual for his or her permission to acknowledge him or her.
