The Use of Digital Images in 24-Hour Recalls May Lead to Less Misestimation of Portion Size Compared with Traditional Interviewer-Administered Recalls^,,

Abstract

Background: The Automated Self-Administered 24-hour (ASA24) dietary recall system enhances the feasibility of collecting high-quality intake data in population-based studies.

Objective: The aim of this study was to assess the accuracy of portion size reporting in the ASA24 compared with interviewer-administered recalls.

Methods: True intake for 3 meals was ascertained in 81 adults aged 20–70 y from the Washington, DC area. Participants were randomly assigned to complete an unannounced ASA24 or an interviewer-administered Automated Multiple-Pass Method (AMPM) recall the following day. An adapted Bland-Altman approach was used to assess agreement between true and reported portion sizes. Linear regression was used to assess log-scale differences between true and reported portion sizes by recall mode. The proportions of reported portion sizes within 10% and 25% of truth were estimated. Analyses were conducted for all foods and drinks and predetermined categories.

Results: Mean differences between true and reported portion sizes were 3.7 g for the ASA24 and 11.8 g for the AMPM. According to the Bland-Altman-type plots, between 92% and 100% (depending on food or drink category and recall mode) of observations fell within the limits of agreement. After adjustment for multiple testing, the mean ratio of reported to true portion sizes was significantly >1 for the categories of all foods and drinks, all foods excluding liquids, amorphous or soft foods, and small pieces among AMPM respondents. Misestimation in the AMPM was significantly different from that in the ASA24 for all foods and drinks and for all foods excluding liquids. Small proportions of reported portions fell within 10% (16.2% for the ASA24 and 14.9% for the AMPM) and 25% (37.5% for the ASA24 and 33.2% for the AMPM) of truth.

Conclusions: The results raise the possibility that digital images tailored to different types and formats of foods may facilitate improved estimation of amounts eaten but highlight the need for continued work in this aspect of dietary assessment. This trial was registered at clinicaltrials.gov as NCT00978406.

Introduction

A fundamental aspect of assessing dietary intake with the use of self-report instruments is portion size estimation (1–4). However, such estimation is challenging and thought to be a substantial contributor to error in dietary intake data (1, 3), including that collected with the use of 24-h recall (24HR)¹¹ methodology. Previous research suggests that portion sizes of different types of foods are reported with differing levels of accuracy. For example, amorphous foods (e.g., pasta or mashed potatoes) may be estimated with less accuracy than single-unit foods (e.g., bagels or cheese cubes) (2, 3), and amounts of foods typically eaten in small quantities (e.g., spreads) may be reported less accurately than other types of foods (5, 6). The existing evidence indicates that, in general, large portion sizes tend to be underestimated and smaller portions overestimated (i.e., flat-slope syndrome) (5, 7, 8). Minimizing misestimation of portion size may be a means of reducing error in self-report dietary intake data (1).

In an effort to enhance accuracy, protocols for administering 24HRs typically incorporate aids to assist participants in estimating the amounts of foods and drinks consumed. The Automated Multiple-Pass Method (AMPM) (9), used in What We Eat in America, the dietary interview component of the NHANES, makes use of 3-dimensional measuring cups and spoons during interviewer-administered recalls and a food model booklet during telephone-administered interviews (10). The booklet includes 2-dimensional photographs of household cups, shapes, and mounds. Estimation of portion sizes for liquids is aided by images of glasses with lines illustrating different amounts.

The development of novel dietary assessment tools that make use of technological advances has allowed tailoring of portion size aids, such as digital images (11–13), to different types of foods. In addition to tailoring to food type, multiple images can be presented, potentially providing respondents with a basis for judging amounts relative to other benchmarks. In comparison with the limited range that may be represented by physical portion size aids (such as measuring cups), the use of multiple images provides the opportunity to present more realistic ranges of amounts of foods and drinks that are typically consumed, which is particularly salient because portion sizes have increased over recent decades (14, 15). Digital images thus have the potential to improve the accuracy with which respondents are able to estimate portion size.

The Automated Self-Administered 24-hour (ASA24) dietary assessment tool is a freely available Web-based tool developed by the US National Cancer Institute that uses multiple passes, adapted from the AMPM (16, 17). The ASA24 system was developed to make it possible to collect multiple recalls in large studies by eliminating the need for an interviewer and implementing automated coding (16). Similar to the AMPM, portion size estimation is a component of the detail pass. However, unlike the AMPM, the ASA24 uses digital images to assist respondents in estimating portion sizes. To assess the accuracy of self-reported intake with the use of ASA24 recalls, a feeding study was conducted in which true intake for 3 meals was documented. We have previously reported on the performance of the ASA24 and AMPM relative to true intake, including the proportion of foods and drinks truly consumed that were reported and concordance between true and reported energy and nutrient and food group intake (18). The differences between reported and true portion sizes for main foods and drinks were 5.4 g for the ASA24 and 14.4 g for the AMPM. For additions to or ingredients in multicomponent foods, the differences between true and reported portion sizes were −0.60 g and 6.2 g for the ASA24 and AMPM, respectively. The objective of the analyses presented here was to conduct a more detailed assessment of the accuracy of portion size reporting in the ASA24 and interviewer-administered AMPM recalls with the use of these same data. We examined the accuracy of portion size estimation for all foods and drinks, as well as for several categories hypothesized to be estimated with varying degrees of error.

Methods

Participants

This study (NCT00978406) included 83 men and women aged 20–70 y who lived in the Washington, DC metropolitan area and were recruited from a database of research volunteers that contained details on sex, age range, and race/ethnicity, allowing for recruitment of a diverse range of participants (18). We sought 40 individuals to complete the ASA24 and 40 to complete the AMPM to enable an analysis of differences in the extent to which respondents accurately reported matches for foods and drinks truly consumed, as reported previously (18).

Eligible subjects had not previously participated in a research study, were not currently dieting, and did not have any formal training in nutrition. Participants were reimbursed for their travel expenses and given modest remuneration for their time. This study was approved by the National Cancer Institute Institutional Review Board, as well as the Westat Institutional Review Board.

Procedures

Feeding, recall, and survey procedures.

The study protocol has been described in detail elsewhere (18). Briefly, data collection was conducted over an 8-wk period in the spring of 2012. Participants were scheduled to visit a study center on 2 consecutive days. On the first day, after being provided with a brief introduction to the study and completing informed consent, they were invited to select and consume foods and beverages from a buffet of breakfast-appropriate items and to return to do the same for lunch and dinner. A variety of food types was offered, both with respect to perceived healthfulness (e.g., fresh fruit compared with brownies) and potential capacity to accurately estimate portion sizes (e.g., amorphous compared with shaped foods) (Table 1).

TABLE 1.

Foods and beverages offered, by category, in a feeding study to compare ASA24 and AMPM interviewer-administered recalls¹

Foods and beverages offered
Liquids
Bottled water
Coffee
Tea
Orange juice
Soda (3 varieties)
Milk
Cream (for coffee)
Salad dressing
Amorphous or soft
Cold cereal (3 varieties)
Oatmeal
Fruit salad
Lettuce (green salad)
Grated cheese (for salad)
Cream cheese (for bagels)
Tuna salad (on sandwiches)
Pesto pasta salad
Rice pilaf
Broccoli
Carrots
Vegetarian lasagna
Sugar (for coffee, cereal)
Single unit
Yogurt (single container)
Chicken breasts and legs
Turkey breast (on sandwiches)
Bread slices (on sandwiches, garlic bread)
Bagels
Apples
Bananas
Potato chips (single-serve bags)
Sugar substitutes (3 varieties, in single-serve packages)
Small pieces
Tomato pieces or slices (in salad and on sandwiches)
Cucumber pieces (in salad)
Red and green peppers (in salad)
Spreads
Margarine
Jelly
Mayonnaise (on turkey sandwich)
Mustard (on turkey sandwich)
Shaped foods
Apple pie (precut)
Chocolate cake (precut)
Brownies (precut)

¹n = 81. AMPM, Automated Multiple-Pass Method; ASA24, Automated Self-Administered 24-hour.

Participants served themselves one at a time at 10-min intervals, and were escorted to a communal dining area to eat. Each container was inconspicuously weighed before and after each participant served him- or herself to determine the amount of each item taken (3, 19). Plate waste was weighed at the conclusion of the meal to enable a calculation of the amount of each food and drink consumed. Weights were taken with the use of Ultra Ship 35 scales, which have a precise accuracy of 0.1 ounces or 2.8 g (≤2 pounds or 0.91 kg) and 0.2 ounces or 5.7 g (>2 pounds or 0.91 kg). Each item was weighed once independently by 2 technicians; if the weights did not match to the gram, a third weight was taken and the mean of the 2 closest weights was used. The weight consumed (i.e., true intake) was calculated as the weight taken minus the weight left (i.e., plate waste). On the second day, participants returned to the center and completed an unannounced 24HR. After stratifying by sex and age group to ensure matching on these characteristics, participants were randomly assigned to 2 groups. One-half completed an ASA24 recall at a computer station and the other one-half completed an AMPM recall administered by a trained interviewer over the telephone. Pairs of participants arrived at intervals across the day so that only one participant was completing either the ASA24 or AMPM at a time. After completion of the recall, participants completed a brief demographic and health behavior survey on a computer at the study center.

Data from 2 participants (one of whom completed the ASA24 and the other of whom completed the AMPM) were excluded because the foods and drinks reported did not correspond to the study center meal offerings, suggesting that they did not report their previous day’s intake. The final analytic sample thus consisted of 81 participants, 40 of whom completed the ASA24 and 41 of whom completed the AMPM. Three of these participants (2 from the ASA24 group and 1 from the AMPM group) did not consume breakfast at the study center and one participant (from the ASA24 group) did not consume dinner there (data for these participants for meals eaten as part of the study were included in analyses). Finally, 2 participants, both in the ASA24 group, did not complete the demographic and health behavior questionnaire.

Portion size assessment.

This study made use of the ASA24–2011 (released September 2011), which included over 10,000 portion size images corresponding to 3800 foods and beverages. The number and format of images presented in the ASA24 were informed by cognitive and usability testing that suggested a preference for multiple simultaneous images (3), consistent with earlier research suggesting the value of a series rather than single photographs to aid in portion size estimation (5). All versions of the ASA24 thus use multiple images, ranging from small to large portions, for each food or beverage. Each image is labeled with the corresponding amount in units appropriate to the given food or beverage, e.g., cups of cereal. The ranges are based on the 5th and 95th percentiles of reported portion sizes from NHANES 2003–2004. For example, images for cereal range from 1/4 cup to 2 cups by increments of 1/4 cup, with options for participants to report amounts less than or greater than the minimum and maximum (Supplemental Figure 1). Foods are photographed on plates or bowls, as appropriate, and framed with cutlery to provide a sense of scale. Images are taken aerially, except for foods for which it is relevant to convey depth (e.g., layer cake); for the latter, images are taken at a 45° angle. For foods such as spreads and condiments that are typically consumed in small amounts, images of household measures (e.g., teaspoons) are used. For amorphous foods such as mashed potatoes, mounds are used. For foods for which size typically varies (e.g., bread), the respondent is first asked to indicate size (e.g., thin or regular) and then to report the amount consumed (e.g., 1 piece or 2 pieces). For beverages, participants are asked to first choose a container type and size and then to indicate the amount actually consumed (Supplemental Figure 2A–C).

Those completing the AMPM recall had access to the USDA Food Model Booklet, measuring cups and spoons, and a ruler to help estimate portion size, simulating an NHANES interview.

Comparison of true and reported intake data.

Data from each recall were reviewed by a nutritionist who was blinded to the true intake data to identify study center meals, with the use of the reported name of the eating occasion, the reported time and location of the eating occasion, and the foods and beverages reported. All foods and drinks reported outside of these meal occasions were excluded from analysis. As a first step in examining concordance between true and reported data, for each food or drink actually consumed, we examined whether the respondent reported a corresponding item, referred to as a “match.” A list of all food codes reported by all participants for the study center meals was generated and assessed by 2 nutritionists to determine whether each was an exact, close, or far match for any of the foods and drinks offered. The identified matches were reviewed by the full study team, and true intake was then compared with reported intake to determine whether each participant reported a match for each of the foods and drinks consumed. Matches were reported for 80% of foods and drinks consumed by those in the ASA24 group and 83% of those in the AMPM group (18). Because an examination of the accuracy of portion size assessment is only possible for foods and drinks for which matches were reported, these matched foods and drinks represent the universe of items considered in the analyses described here.

Before analysis, corrections were applied to address known errors in the ASA24–2011 database (20) and to address ASA24 system-related issues that have since been remedied (affecting 2% of foods and drinks reported in the ASA24 in this study). Furthermore, a review of free text entered by ASA24 participants who selected “other” as a response to a detail question or who indicated that they could not find the food or drink that they consumed (i.e., “match not found”) was conducted by 2 trained coders and corrections were made if necessary (affecting 1% of foods and drinks reported with the use of the ASA24). A more extensive analysis of the implications of editing free text has shown that such editing does not have a large impact on nutrient and food group estimates (21); however, for the purposes of this validation study, corrections were made to ensure maximum accuracy when comparing true and reported intake.

Statistical analyses

Analyses were conducted with the use of SAS, version 9.2.

Previous research suggests that there is wide variation in accuracy of portion size estimation across foods and food types (2). We thus conducted analyses for all foods and drinks combined, as well as predetermined categories [all foods excluding liquids, amorphous/soft foods, single-unit foods, small pieces, shaped foods, and spreads (Table 1)] on the basis of potential differences in ability to accurately estimate amounts of items with these different characteristics.

A common approach for assessing agreement between imperfect measures is the Bland-Altman method (22). Because we documented true intake, we adapted the Bland-Altman method to plot the differences between reported and true portion size estimates against truth (rather than against a mean of values from 2 error-prone measures). The resulting Bland-Altman-style plots suggested that the variability of reporting errors increased with true portion size. To account for this, we plotted differences between log-transformed values of reported and true portion sizes (equivalently, the log of the ratio of reported to true portion sizes) against the log of true portion size values. These plots (which suggested less heteroscedasticity, or weaker relations between error variability and true portion size) were used to assess the limits of agreement between true and reported portion sizes for each of the ASA24s and AMPMs. We back-transformed the mean difference in the log scale to obtain an estimated mean ratio of reported to true portion sizes separately by recall mode. To assess whether agreement between true and reported portion sizes differed by recall mode, we tested the coefficient of the mode indicator in a linear regression model fit to the differences in log-transformed values.

To provide further insights into the implications of the accuracy of portion size estimation, reported portion sizes were determined to be within a certain range of truth. Because there does not appear to be a consensus in terms of what range of accuracy is acceptable (2), we examined proportions within 10% and 25% of truth and used logistic regression to examine differences in the odds of meeting these criteria by recall mode, for all foods and drinks and for the predefined categories.

We previously reported that the ASA24 and AMPM groups in this study differed in terms of education, race/ethnicity, and use of vitamin and mineral supplements (18). Multivariate analysis showed that after accounting for race/ethnicity, the differences in the other characteristics were not significant; thus, race/ethnicity was included as a covariate in regression models that compared the 2 groups. To account for the possibility that each individual could contribute multiple potentially correlated observations to analyses, SEs were estimated with the delete-one jackknife procedure applied at the individual (rather than the observation) level. Thus, the effective degrees of freedom for mean estimates were often substantially smaller than the raw numbers of observations comprising the mean. We used the Benjamini-Hochberg step-up procedure (23) with a false discovery rate of 5% to adjust for multiple comparisons. This was done separately for the tests of differences between true and reported portion sizes within recall modes and between recall modes.

Because some extreme values of true and reported intake were apparent in the data, we conducted sensitivity analyses to assess the influence of outliers. The data for all foods and drinks were first transformed to approximate normality, values outside of 3 times the IQR were removed, and the analyses were repeated (both for all foods and drinks and for the predetermined categories). Excluding these outliers had little impact on the nature of our findings; thus, we present the results with all data points included.

Results

The characteristics of the study sample by recall mode (ASA24 compared with AMPM) have been described previously (18). The 2 groups were well matched on sex and age range as per the study design. As noted, the groups differed on race/ethnicity (ASA24: 67.5% white; AMPM: 34.1% white), and this factor thus was accounted for in comparisons between the groups.

The spread of the untransformed true and reported portion sizes by food or drink category for the ASA24 and AMPM groups, respectively, is illustrated in Supplemental Figures 3 and 4. The Bland-Altman–style plots for all foods and drinks for the ASA24 and AMPM are shown in Figures 1 and 2. Typically, when the Bland-Altman method is used, measures are considered comparable if >95% of data plots fall within 2 SDs of the mean. The plots indicate that, overall, 94% of log-scale differences fell within these limits of agreement for both the ASA24 and AMPM. Plots differentiated by category of foods and drinks are available in Supplemental Figures 5–18. The fraction of observations that fell within the limits of agreement ranged from 92% to 100% for the ASA24 and 94% to 97% for the AMPM across categories.

FIGURE 1.

Agreement between log(Amount eaten, g) and the log of the ratio between reported to true portion sizes for all foods and drinks for which matches were reported, based on Automated Self-Administered 24-hour dietary assessment tool respondents (n = 40 individuals and 801 observations).

FIGURE 2.

Agreement between log(Amount eaten, g) and the log of the ratio between reported to true portion sizes for all foods and drinks for which matches were reported, based on Automated Multiple-Pass Method respondents (n = 41 individuals and 900 observations).

Mean true and reported portion sizes for all foods and drinks and for the food and drink categories for the ASA24 and AMPM respondents are provided in Table 2. Mean differences are provided for illustrative purposes; for all foods and drinks for which a match was reported, the mean differences between true and reported portion size were 3.7 g for the ASA24 and 11.8 g for the AMPM. These differences equate to 3.6% and 13.1% of the true mean portion size for the ASA24 and AMPM, respectively. The mean ratio of reported to true portion sizes was significantly higher than one for all foods and drinks, all foods excluding liquids, amorphous or soft foods, and small pieces among the AMPM respondents. For the ASA24, the nominal P value for spreads was <0.05 (indicative of underestimation), but this difference was not judged to be significant with adjustment for multiple testing. In terms of differences by mode of recalls, the mean ratio of reported to true portion sizes was significantly different between the ASA24 and AMPM for all foods and drinks and for all foods excluding liquids, with a tendency toward overestimation in the AMPM.

TABLE 2.

Mean true and reported portion sizes and differences, mean ratio of amount reported to amount eaten, and proportion of reported portion sizes within 10% and 25% of truth for foods and drinks for which a match was reported, by recall mode, ASA24 and AMPM¹

	ASA24							AMPM
	Observations, n	Mean amount eaten, g	Mean amount reported, g	Difference,2 g	Ratio of amount reported to amount eaten	Within 10% of truth, %	Within 25% of truth, %	Observations, n	Mean amount eaten, g	Mean amount reported, g	Difference,2 g	Ratio of amount reported to amount eaten	Within 10% of truth, %	Within 25% of truth, %
All foods and drinks	801	104	108	3.68 (−3.76, 11.1)	0.977 (0.896, 1.06)	16.2	37.5	900	90	102	11.8 (5.75, 17.9)	1.17 (1.09, 1.25)	14.9	33.2
All foods excluding liquids	637	62.9	69.8	6.85 (0.471, 13.2)	0.98 (0.892, 1.08)	14.4	35.5	731	60.1	73.4	13.4 (7.99, 18.8)	1.21 (1.12, 1.3)	15.0	32.7
Liquids	164	264	255	−8.66 (−33.7, 16.4)	0.963 (0.839, 1.10)	23.2	45.1	169	219	225	5.13 (−14.7, 24.9)	1.00 (0.902, 1.11)	14.2	35.5
Amorphous or soft	303	66.1	80.5	14.4 (5.07, 23.8)	0.964 (0.859, 1.08)	7.92	28.4	338	62.8	83.7	20.9 (12.5, 29.3)	1.20 (1.09, 1.33)	12.4	32.8
Single unit	192	74.5	73.1	−1.42 (−8.26, 5.43)	1.02 (0.929, 1.12)	26.6	53.1	209	72.4	70.9	−1.48 (−6.03, 3.06)	1.04 (0.956, 1.13)	22.5	44.0
Small pieces	60	17.9	21.8	3.89 (−3.22, 11)	1.17 (0.837, 1.65)	6.67	16.7	83	18.3	40.1	21.8 (11.7, 31.9)	1.97 (1.61, 2.41)	8.43	14.5
Shaped foods	47	94.3	95.5	1.22 (−15.7, 18.1)	0.887 (0.753, 1.05)	17.0	34.0	65	85.9	103	17.4 (0.989, 33.8)	1.08 (0.856, 1.36)	13.8	20.0
Spreads	35	7.61	6.97	−0.639 (−2.36, 1.08)	0.764 (0.596, 0.98)	14.3	34.3	36	12.1	14.7	2.59 (−2.38, 7.57)	1.24 (0.918, 1.67)	13.9	30.6

¹Values are means (95% CIs) unless otherwise indicated. ASA24, n = 40 individuals and 801 observations; AMPM, n = 41 individuals and 900 observations. AMPM, Automated Multiple-Pass Method; ASA24, Automated Self-Administered 24-hour.

²Expressed as reported minus true portion size. Thus, positive differences are indicative of overestimation and negative differences are indicative of underestimation of portion size.

The proportion of foods and drinks for which the reported portion sizes were within 10% and 25% of true portion sizes (in grams) is also shown in Table 2. For all foods and drinks, the proportions within 10% of truth were 16.2% for the ASA24 and 14.9% for the AMPM; the proportions within 25% of truth were 37.5% and 33.2% for the ASA24 and AMPM, respectively. In each case, the lowest proportions were observed for small pieces and the largest proportions were observed for single-unit foods. No significant differences in the proportions within 10% and 25% of truth by recall mode were observed.

The mean differences between true and reported portion sizes and the mean ratio of reported to true portion sizes for each food and drink individually are provided in Supplemental Table 1. Jackknife CIs for some individual foods and drinks may be unreliable because of a combination of small frequencies of consumption and small numbers of distinct respondents comprising the mean.

Discussion

The results of this study suggest that accuracy of portion size reporting differs between respondents completing self-administered 24HRs with the use of the ASA24, which relies on digital images to facilitate estimation of portions, and those completing interviewer-administered recalls with the use of traditional portion-size aids. Among AMPM respondents, the mean ratio of reported to true portion sizes was significantly greater than 1, indicating overestimation, for the categories all foods and drinks, all foods excluding liquids, amorphous or soft foods, and small pieces. Furthermore, misestimation in the AMPM was significantly different from that in the ASA24 for all foods and drinks and all foods excluding liquids, with overestimation apparent in AMPM respondents. These findings raise the possibility that images tailored to different types and formats of foods may facilitate improved estimation compared with static images and household measures. Nonetheless, further enhancements to enable accurate portion size estimation are needed, as is evidenced by the proportions of reported portion sizes that were within 10% and 25% of truth: among ASA24 respondents, only 16% and 38% of estimated portion sizes were within 10% and 25% of the amounts eaten, and among AMPM respondents, these figures were 15% and 33%, respectively.

Previous research suggests that underreporting of large portion sizes may be pronounced when the portion size aids are not well matched to the size of typical portions (8). Commonly used portion size aids may not reflect the growth in portion sizes that has occurred in recent decades (14, 15). The use of multiple images covering a range of serving sizes derived from food consumption data may address this challenge to some extent by providing a more realistic universe of possible portion sizes from which participants can select. Earlier formative research to guide ASA24 development suggested that a greater number of photos (8 compared with 4) might be beneficial in terms of accuracy (3), consistent with a previous study (5). Nonetheless, despite the use of a range of images, the patterns in the data suggested that true portion size was associated with variability in misreporting.

In addition, portion size reporting previously has been shown to be especially challenging for particular types of foods (2, 3, 5, 6). We similarly found variation in accuracy across different categories of foods. The lowest proportions of reported portion sizes within 10% and 25% of truth were observed for small pieces and the largest proportions were observed for single-unit foods. For the AMPM, portions were overestimated for amorphous or soft foods, as well as those offered as small pieces. We did not find that portions were significantly misestimated for liquids, indicating that they did not drive differences observed for all foods and drinks in the AMPM, despite the fact that they tended to be consumed in larger quantities. It should be noted that comparisons across food categories are limited because of a small number of foods within categories, as well as clustering of types of foods within some categories. For example, the 3 items categorized as “shaped foods” were each desserts that had been precut into squares or wedges (although there are differences in the portion size depictions in the ASA24, because angled images are used for cake to provide a sense of depth, whereas aerial images are used for pie and brownies). In this case, not only was the universe of choices limited, but it is possible that there is a particular bias in the reporting of these foods (potentially perceived as less healthy by respondents in comparison with some other foods offered, such as fresh fruit), including the amount consumed. Furthermore, the determination of categories for analysis was a judgment call among multiple nutritionists, and some of the food offerings could have been considered to fit in other categories. However, these categories were identified a priori on the basis of previous research.

Our findings suggest that, consistent with other research, both under- and overestimation are possible with portion size reporting (2). Interestingly, although we observed overestimation in the AMPM, the ratios of reported to true portion sizes suggest underestimation in the ASA24, although formal significance was not attained. In our previous analyses of these data, we found reasonable agreement between estimated intake of energy, nutrients, and food groups on the basis of mean true and reported consumption, despite the fact that ∼20% of foods and drinks consumed were not reported (18). Many of the items excluded from reported intake were additions to foods, such as tomatoes in salad or lettuce on sandwiches that are generally consumed in small amounts, with perhaps little impact on intake estimates. Exclusions also may be compensated for by intrusions (foods reported but not consumed). Another potential contributor to the consistency between true and reported nutrient and food group estimates, despite excluded foods and beverages, may be overreporting of portions consumed for some items. As an example, the portion size of regular soft drinks was overestimated by AMPM respondents by 24 g on average. This would result in an overestimation of 9 kcal and 2 g sugar (24). Lasagna was overestimated on average by 21 g by ASA24 respondents, which corresponds to 27 kcal, 1.5 g protein, 25 mg calcium, and 8 mcg folate (24). These errors are compounded over all foods and drinks consumed when considering total intake. However, the estimates are averaged over a wide range of potential magnitudes of misestimation at the individual level; variability across individuals as indicated by the CIs indicates that effects on estimated energy and nutrient can be much smaller or larger.

Since this study was conducted, a mobile-friendly version of the ASA24 has been released. Unlike previous iterations, this version (ASA24–2016) is functional on tablets and smart phones, which necessitated changes to the way in which portion sizes are presented. Although multiple images continue to be presented for each food and drink, a limited number of images is visible at one time, depending on the size of the device being used. The respondent can scroll through the images, preserving the capacity to view and compare images representing a range of portion sizes. An earlier small study provided some evidence that the accuracy of portion size estimation is better with simultaneous presentation of images (3). Further research to test this and other ways of capturing portion size in dietary assessment are thus warranted. Indeed, continued technological advances in the capture of dietary intake data (11–13), including the use of smart phones to capture images before and after eating combined with fiducial markers that provide an aid to automated volume estimation, have the potential to enhance the accuracy with which amounts of food consumed can be captured. Such advances may be accompanied by disadvantages, such as the reactivity that is inherent with real-time recording (4, 25), as well as lapses in data collection (e.g., missing images or difficulty identifying the foods photographed and their portions). As a result, it is imperative to continue to weigh the pros and cons of different methods of data capture, including portion sizes, for different study types and research objectives (4, 25). This extends to the use of the ASA24 to collect food records, a capability added in the latest version; the validity of ASA24 records has not yet been evaluated.

This study is not without limitations. It was powered primarily to allow a comparison of the proportion of foods and drinks consumed for which a match was reported between the ASA24 and AMPM groups, and the statistical power for an examination of true compared with reported portion sizes was limited. The offerings and frequency of consumption and corresponding reporting of matches for items within the categories examined were also limited. Although we took pains to ensure that the findings judged to be significant were likely to be real, the small sample size makes it possible that we could have missed important differences in both instruments. The sample consisted of volunteers who received remuneration and may have been particularly motivated to accurately report what they consumed. Furthermore, the study included adults aged 20–70 y, and the findings may not be generalizable to children or older adults. Factors thought to influence the accuracy of portion size reports with the use of digital images include the ability to recall the amount of food eaten (memory), to relate a food amount actually present to a portion size aid (perception), and to develop a mental picture of a food portion not present and relate it to a portion size aid (conceptualization) (5). The potential for these factors to vary by factors such as age suggests the need for evaluative work with other populations, such as children. The results also may not be generalizable to those with low literacy and/or numeracy skills (26). In addition, the ASA24 requires access to high-speed internet, which may limit its use in certain circumstances, although this is likely to be less of an issue moving forward with continued advances in technology. A large community-based study conducted in adults with known Internet access showed that participants had high completion rates of 2 nonconsecutive recalls with the use of either interviewer-administered AMPM or ASA24 recalls, and that respondents overwhelmingly preferred the ASA24 to interviewer-administered recalls (27). This suggests that the ASA24 may be viable for use in a range of studies.

Our previous analyses indicated similar accuracy in terms of the proportions of foods and drinks actually consumed that were reported and corresponding nutrient estimates between respondents completing recalls with the use of the ASA24 and AMPM. The current results provide further evidence that the ASA24 performs well relative to true intake and to interviewer-administered recalls. In sum, the ASA24 provides a viable alternative to interviewer administration for the collection of 24HR data. Nonetheless, our results also highlight challenges to the accurate estimation of portion size, speaking to the need for continued work in this area as a means of improving self-reported dietary intake data and our capacity to understand diet–health relations. Based on their assessment of error in portion size estimation with the use of computer-based portion anchors, Hernández et al. (2) concluded that “precision in portion size estimation is not yet a realistic expectation.” Our results suggest that this remains true, at least on the basis of the use of digital images as portion size aids.