Given the dearth of detailed, periodic dietary intake data for much of the world’s population and the volume of food consumption data that is present in household consumption and expenditure (HCE) surveys, the potential value of HCE data to nutrition research and surveillance is immense, particularly for developing countries. In recognition of this, recent decades have seen steadily growing interest in survey design and analytical approaches geared toward increasing the applicability of HCE data in nutrition [7
]. This effort is challenged by the fact that household food and nutrient consumption are far from perfect proxies for individuals’ diets, the primary exposure of interest in nutritional epidemiology and one that remains difficult to assess with great accuracy even under the best of conditions. Nonetheless, necessity is the mother of invention, and some interesting ways to meet this challenge have suggested themselves in the literature, four of which are evaluated in this paper.
4.1. Aim 1: Direct Comparison between Per-Capita Household Consumption and Per-Capita Dietary Measurements from the Same Households
The first and simplest approach involves direct inference of dietary intake based on per-capita household food consumption [32
]. Accurate household consumption measurements are a prerequisite for accuracy of the AME and statistical disaggregation methods that are evaluated in this paper; for the purpose of directly assigning dietary intake to individuals, per-capita household consumption measurements are less useful for multi-person households because they imply impossibly equitable intra-household distribution of food. Because persons living by themselves are the main consumers of food in their household (with the exception of guest and visitors), household food consumption may be an appropriate proxy for these individuals’ dietary intakes, although the degree to which these estimates are generalizable to those living in multi-person households may be limited. In the current study, we found household food group and nutrient consumption among 109 FCS-HH households fully-enumerated by the FCS-24 to overestimate and correlate poorly with dietary intake in both types of households, especially for single-person households.
Important sources of systematic and random error are known to influence reporting of household food consumption data [10
] and have more recently been subject to more formal decomposition [36
]. In particular, the magnitude of overestimation in the FCS-HH suggests that enumerators provided telescoped estimates (their recall included household foods consumed prior to the reference period) [36
]. It is also plausible that reported household food consumption was partly conflated with food that was acquired (or simply present in household stocks) over the reference period, but that not necessarily consumed, or was perhaps transferred to other households. It is not immediately clear why over-reporting would affect multi-person households to a lesser extent than single-person households, but this may have to do with accuracy that is incurred by the cognitive exercise of distinguishing and dividing consumption among each household member in a multi-person household, while those living alone might rely on less enumerative rules-of-thumb. Recall error is mitigated by the use of prospective instruments such as the HSES-HH’s consumption diary, but these are conversely more likely to be affected by underreporting due to respondent burden [21
]. Both retrospective and prospective assessment methods are also subject to some measure of social desirability bias (the latter by way of respondent “self-monitoring”) [38
], however, research is lacking on the importance of this bias in assessing diet and household food consumption in Mongolia. Efforts to improve the accuracy of household food consumption measurements are ongoing and have considered such cognitive and survey design [7
]. In Mongolia, recent analysis by Troubat and colleagues suggests that the HSES-HH’s diary instrument could be satisfactorily substituted with a less costly consumption recall combined with measurement of changes in household foods stocks and acquisitions [21
Per-capita estimates of household consumption density more closely agreed with dietary-derived per-capita intake densities in both household types, particularly in the case of nutrients. To some degree, this agreement may be inflated by shared systematic error in the reporting of foods in the FCS-HH and FCS-24 (e.g., the fact that both rely on memory and self-report), as well as nutritional analysis (e.g., the fact that the same food composition data was used to analyze both datasets) [40
]. The observed agreement is nonetheless encouraging given that nutrient densities are meaningful nutrition indicators in and of themselves, and which provide a convenient way to compare individuals with different caloric intakes despite the aforementioned sources of error (both of which are non-differential) [28
An important area of innovation in dietary and household surveys globally is the emerging use of technology-assisted assessment of food consumption, as well as commercially-available retail and procurement data from supermarket sales and purchases [41
]. To our knowledge, technology-assisted collection methods are not widely used in Mongolian national surveys, including in urban areas. Since 2016, the Mongolian Tax Administration has collected extensive information about individuals’ purchases of specific goods and services in Mongolia using an advanced electronic database system [44
] for the purpose of issuing tax refunds, however, it is not clear how applicable this information may be to the analysis of food consumption patterns.
4.2. Aims 2 and 3: Comparison between Disaggregated Household Consumption Estimates and Individual Dietary Intake Measurements
Next, we evaluated the validity of the AME method to disaggregate household food and nutrient consumption based on household members’ relative caloric requirements. The validity of this method has been evaluated in numerous surveys outside Mongolia. In studies of two household consumption and expenditure surveys in Uganda, the AME method provided reasonable estimates of dietary nutrient density, but more often underestimated the dietary intake of potential fortification vehicles among women and children in comparison with results of a nested 24-h recall, varyingly explained by the inability of each survey’s household instrument to fully enumerate-foods consumed and the extent to which the intra-household distribution of staple foods in Uganda is disproportionate to the caloric requirements of household members [46
]. By contrast, an analysis of 4195 Bangladeshi households revealed the AME method to produce remarkably accurate disaggregated estimates of most nutrients’ dietary intake in comparison with the results of 24-h recalls collected from the same study population, implying that the consumption of most foods would likely be accurately disaggregated as well [48
]. This finding was corroborated by a pooled analysis of six Bangladeshi surveys including 1232 households, which found that in Bangladesh, more so than in most of the 13 other countries for whom similar pooled analyses were undertaken, intra-household distribution of consumed calories appears to be relatively proportional to intra-household distribution of caloric requirements (this is a necessary, though not sufficient prerequisite for intra-household distribution of foods and non-caloric nutrients in a manner that is proportional to caloric requirements, which is a cardinal assumption of the AME method) [16
In the present study, the application of the AME method to the larger HSES-HH showed it to be generally more apt than the statistical method at estimating and ranking individuals’ intakes of dietary components, but it also overestimated intake in both household surveys and produced extremely narrow standard errors. The latter may be attributed to the AME method’s relatively deterministic manner of disaggregating consumption, which could be addressed by assigning more granular estimates of energy expenditure (or by deliberately assigning error to estimates, drawn from error observed in energy expenditure prediction models [49
]). On the other hand, a benefit of a deterministic approach is that it does not imply a sample size requirement to produce precise disaggregated estimates (unlike the statistical method). With regard to the comparative accuracy of the AME method, some investigators specifically suggest that its strength lies in estimating the intake of those dietary components more correlated with energy [50
]. Accordingly, in disaggregation of the HSES-HH, the AME method more accurately estimated individuals’ intakes of animal fat/eggs/dairy products, baked and fried flour products, and flours/grains/noodles, which are the major staples of the Mongolian diet, and which are relatively calorie-dense and nutrient sparse. To the extent that dietary intake of caloric energy, macronutrients, and staple foods (for example, fortifiable flour) are ubiquitous and are subject to homeostatic regulation [52
], predicting individuals’ intakes of these dietary components should require a disaggregation method to be less discriminating of components of variation in intra-household food consumption, which are attributable to prevailing social or cultural forces rather than biological ones. In such cases, it may be more reasonable to depend on the AME method than the statistical method, the latter of which may incur statistical error without a discernible benefit to accuracy. The AME method may also be extended to a more generalized concept of intra-household “equivalency scales” by weighting nutrient household consumption according to nutrient requirements other than that of energy [50
]. If household food consumption is reported inaccurately, however (as verified in the case of the FCS), the AME method will produce biased estimates regardless of dietary components’ known associations with energy or other nutrients’ intake or requirements.
Unlike the AME disaggregation method, the statistical method has not been previously validated. The plausibility of dietary intake estimates that are produced by the statistical method generally supported in the literature by its apparent ability to predict natural variation in caloric intake with age—increasing intake during childhood, a spike in puberty, and a decrease later in life—rather than by comparison with actual consumption data for energy or other dietary components (which had not previously been studied) [18
]. Despite this, an advantage of a more data-driven statistical method over that of the AME would be expected in the case of dietary components, whose consumption is less correlated with energy requirements (e.g., most foods (Table S4
)), and those that are less correlated by definition (all food group and nutrient intake densities). Accordingly, the statistical method more accurately assigned dietary intake of food groups and intake densities of both food groups and nutrients in disaggregating both household surveys.
An interesting aspect of the statistical method is that its inclusion of a model intercept accommodates the possibility that not all household food is consumed, and thus ought to be disaggregated, which could be important if household consumption were measured in terms of proxies, such as food expenditure, acquisitions, or stocks; this is suggested by Chesher in the method’s initial application to food acquisitions among British households [18
]. In disaggregating surveys, which explicitly measure household food consumption (such as those analyzed in this study), the intercept explicitly represents consumption that is unrelated to the number, age, or sex of individuals living in each household, which may be useful if it helps account for food that was reported to be consumed, but which was in fact merely acquired, present in the house, but not consumed, given to animals, wasted, or which spoiled. This usefulness is supported by the statistical method’s comparative accuracy in disaggregating consumption in the FCS-HH (Aim 2) despite this survey’s overestimation of per-capita dietary intake (Aim 1). The utility of the intercept in this regard requires that household consumption is over-reported in an additive rather than a multiplicative fashion, otherwise the differences between predicted intakes across age-sex groups will be inflated (as will the model intercept); we have affirmed this experimentally by applying the statistical method after adjusting household consumption using either a constant or a multiplier (not shown). Conversely, to the extent that household food consumption is multiplicatively underestimated, the intercept will be attenuated, as will the differences in predicted intake across age-sex groups. This may have been responsible for the statistical method’s poor performance in the disaggregation of the HSES-HH (which likely experienced multiplicative underreporting associated with the burden of the diary instrument), and why performance improved after removing the model intercept in applying the AME-like (“SD2”) adjustment. Thus, while the statistical method depends less on assumptions of accurate reporting of household consumption per se than the AME method, it is nonetheless influenced by the nature of this inaccuracy.
The statistical method is potentially limited in ways that the AME method is not, stemming from its reliance on accurate and precise prediction of household food consumption (without which accurate or precise estimates of dietary intake among different age-sex groups may not be inferred). For example, zero-inflation in the distribution of household food consumption due to the presence of non-consumers over the reference period may produce poor model fit and inaccurate predictions (Table S2
). In this study, our use of zero-inflated models implies that non-consuming households would in fact be consumers given a longer reference period, which is likely a reasonable assumption for most food groups and nutrients (except alcohol), but that may not be reasonable were smaller (less aggregated) food groups to be analyzed. In such cases, a two-part or “hurdle” model which deliberately distinguishes between processes of household consumption frequency and consumption magnitude may be more appropriate for modeling mean household consumption in the population. With regard to precision, while smoothing parameter estimates may be helpful for producing more realistic estimates, the degree of smoothing is a subjective choice that may obscure rather than expose true variation in predicted dietary intake with age, particularly if the imprecision in estimates is severe. In this study, model fit of statistical disaggregation models was generally poor (Table S6
). Improving precision is challenged by the fact the inclusion of highly predictive variables—household energy intake or household size—e.g., changes the interpretation of parameter estimates, such that they reflect effects on household composition rather than the addition of household members (partly defeating the purpose of using the statistical method over the AME method, the latter of which is necessarily dependent upon assumptions of intra-household distribution).
4.3. Aim 4: Direct Prediction of Dietary Nutrient Intake by Individuals
Finally, we attempted to estimate individuals’ dietary intakes and intake densities using a prediction model that was incorporating household food consumption and other data feasibly obtainable from a household survey, with relatively precise results. While examples of this approach are relatively sparse in the literature [51
], we derived what we consider to be acceptably precise predictions of dietary nutrient intake and intake densities. Given the predictors that are available for model selection, the results were similar between models directly predicting nutrient intake densities vs. those based on separate prediction of nutrient and energy intake. Similar to the statistical disaggregation method, the prediction model does not require potentially inaccurate assumptions about the intra-household distribution of food consumption. Prediction further relaxes assumptions that the reporting of household food consumption is systematically (as in the case of the AME method) or differentially (as in the case of the statistical method) unbiased with respect to dietary intakes of household members, and offers more flexibility with respect to potential effect modifiers or confounders. For example, we found in Aim 1 that bias in per-capita household consumption was differentially affected by household size. In Aims 2 and 3, we found that despite attempting to control for household educational attainment, family composition, outside food spending, consumption by impermanent members, and locality, a strong pattern of increasing estimated intake in advanced age was observed in both the AME and statistical disaggregation estimates for most foods and nutrients, which is contrary to what we expected based on both dietary energy intake and predicted energy expenditure. This pattern may result from residual confounding by socioeconomic status and household size, in that wealthier Mongolian households generally consume more food, are smaller (increasing per-capita food consumption), and their members have longer life expectancies [60
]; it is also possible that smaller (and younger) households underreported food consumption, according to the cognitive hypothesis that was discussed previously in Aim 1 (to some extent, increasing intake with age may also be real, given that the Mongolian population is still relatively young and older individuals are more metabolically active than their counterparts in other populations). In addition to more efficient control for confounding variables, prediction allows for estimates to be produced across more granular strata of individuals, while the statistical method may only do so with difficulty (for example, by analyzing strata independently and reducing statistical power, or by introducing a potentially unwieldy number of interaction terms between age-sex groups and covariates of interest [18
Based on our prediction models, we suggest that household surveys would be well adapted to estimate dietary intake and intake densities by the addition of a rudimentary dietary assessment module. Predictive approaches have performed well in analysis of food frequency questionnaires (FFQs), the rationale being that such an approach acknowledges “the importance of a food item should reflect not only the nutrient content of the food, but also the validity of the responses to that particular item” [37
]. Existing platforms for conducting household surveys would be well-suited for applying this method, given that they are prepared using large, nationally-representative sample frames, and are collected periodically. The expense of a validation study (i.e., simultaneous collection of dietary intake data with which to build a model) should not be considered as a limiting factor, as it will also produce useful consumption estimates that could otherwise have been collected in a separate dietary survey; even cursory qualitative information about individuals’ diets can be useful for assessing food security or screening for chronic disease risk [6
]. Still, some may question the purpose of adding dietary assessment of individuals to a household survey in lieu of conducting a more rigorous standalone dietary assessment. If resources were available to do so, then measurements that are collected in such a survey would assuredly be more accurate than those that were obtained through prediction. If resources are not available, however, prediction may offer a reasonable compromise between an infeasible approach and no approach at all. Furthermore, while it is not unreasonable to append a qualitative or semiquantitative dietary assessment module to an HSES questionnaire, more involved dietary measurements (such as diet records or a 24HR) may diminish compliance and compromise accurate collection of other survey modules. For the purpose of prediction, the level of detail at which to collect individuals’ dietary and eating behavior information—cursory qualitative, cursory semiquantitative, or detailed semiquantitative—should be carefully considered in the context of a given HCE platform, not all of which may be suited to accommodate a highly detailed questionnaire. This should not preclude the consideration of more detailed quantitative or semiquantitative food frequency questionnaires, however (the value of which could not be evaluated in this study given the use of the 24HR). If an FFQ were used, a predictive model framework in the context of HCE data may enhance the instrument’s usefulness in collecting absolute intake, while in the case of a 24HR, it may increase the instrument’s ability to assess long-term diet.
Finally, household consumption data and household characteristics may also provide added value when being collected as part of a dietary survey or otherwise in the context of national dietary data (rather than simply as a less expensive alternative to dietary data) by supplying complimentary information on the intra-household distribution of food consumption and nutrient adequacy, and household-level predictors of food security and dietary diversity [5
], which would not generally be assessed in a dietary survey. Furthermore, while dietary surveys are collected using a variety of methodologies, efforts to standardize household survey instruments have rendered household consumption data generally more appropriate for international comparisons, and these data may therefore be generally more applicable to multilateral policymaking [30
4.4. Strengths and Limitations
By disaggregating two household surveys from the same national population using two different instruments for assessing household food consumption (a recall and a diary), this study was able to assess the reproducibility of disaggregated household consumption estimates and study differences in the survey design. The size of the HSES-HH allowed for more statistically powerful disaggregation, while the FCS-HH, although a smaller survey, was conducted in the same population as the dietary assessment, and thus allowed for an inherently more direct and multi-faceted comparison. Comparability of the two household surveys was strengthened in that both were nationally-representative, seasonally-matched, and conducted within two years of one another. Analysis of both individual dietary intake and household food consumption incorporated local and empirical food yield, food composition, and physical activity, and incorporated empirical estimates of food eaten outside of the home, allowing for more a more rigorous validation.
An important limitation of this study is potential underreporting by the 24-h recall. The extent to which this has affected the comparative validity of the AME and statistical disaggregation methods is expectedly mitigated inasmuch as this underreporting affected both household surveys in a similar fashion, the fact that all of the disaggregated household results were compared to the same dietary assessment, the fact that dietary underreporting should not necessarily be expected to differentially bias reported or predicted intake of a given food group or nutrient across different age and sex groups; in validating both per-capita estimates (Aim 1) and disaggregated estimates (Aims 2 and 3), underreporting was further mitigated by conducting energy-adjusted analysis [28
]. Second, while being suitable for assessing mean dietary intake, a single 24HR does not provide estimates of usual intake. Adjustment for within-person variance using variance components from the same national population helped to account for this limitation in the case of nutrients, but not food groups. Third, while various factors were applied to render household food consumption measurements comparable with dietary intake, including consumption by impermanent household members, we were unable to account for guests and visitors who affect household food supplies, but were not accounted for by the surveys that were analyzed. A final limitation of this study was the lack of information on individual dietary intake by children, or by any age groups in seasons other than summer, making it impossible to determine the validity of the method for children or in different seasons in Mongolia. Further research is warranted to address this.