Predictive equation derived from 6,497 doubly labelled water measurements enables the detection of erroneous self-reported energy intake

Table of Contents

Overall Summary

Study Background and Main Findings

This study developed a predictive equation for Total Energy Expenditure (TEE) using the International Atomic Energy Agency Doubly Labeled Water Database, encompassing 6,497 measures from individuals aged 4 to 96. The equation, derived via general linear regression, showed strong predictive power, with 94.6% of independent TEE measurements falling within the 95% predictive limits. Application to the NDNS and NHANES datasets revealed significant misreporting, with over 50% of dietary reports falling outside the predicted TEE limits. Notably, under-reporting was strongly correlated with higher BMI and reported protein intake, while being negatively correlated with reported fat intake.

Research Impact and Future Directions

The study provides a robust predictive equation for TEE and offers valuable insights into the pervasive issue of misreporting in dietary surveys. The strong correlation between under-reporting and factors like BMI and reported macronutrient intake highlights the limitations of relying solely on self-reported dietary data. However, the study clearly distinguishes between correlation and causation, acknowledging that the observed relationships do not prove a causal link between dietary composition and misreporting.

The practical utility of the predictive equation is significant, as it offers a more accurate tool for identifying potential misreporting in large-scale dietary surveys. This can help improve the accuracy of nutritional epidemiology research and inform the development of more effective public health interventions. The findings are placed within the context of existing literature, building upon previous work on misreporting and offering a novel approach based on a large and diverse dataset.

While the study provides valuable guidance for researchers and practitioners, it also acknowledges key uncertainties. The reliance on self-reported data, even with the use of a predictive equation, remains a limitation. The study also highlights the need for caution when interpreting self-reported dietary data, particularly in relation to macronutrient composition and its association with BMI. The authors suggest that future improvements may involve integrating objective measures of physical activity, such as accelerometry.

Critical unanswered questions remain, particularly regarding the generalizability of the findings to populations not represented in the study sample. Additionally, while the study identifies patterns of misreporting, the underlying reasons for these patterns are not fully explored. Further research is needed to investigate the psychological and social factors that contribute to misreporting. The methodological limitations, such as the exclusion of children under 4 and the handling of 'other' and mixed-race ethnicities, do not fundamentally affect the main conclusions but highlight areas for future research. Overall, the study makes a significant contribution to the field of nutritional epidemiology by providing a valuable tool for identifying misreporting and highlighting the need for more accurate methods of dietary assessment.

Critical Analysis and Recommendations

Comprehensive Data Utilization (written-content)
The study leverages a large and diverse dataset from the International Atomic Energy Agency Doubly Labeled Water Database, enhancing the robustness and generalizability of the predictive equation. This is crucial because it allows for a more accurate representation of the population and increases the applicability of the findings.
Section: Abstract
Detailed Analysis of Survey Data (written-content)
The application of the predictive equation to two large dietary surveys (NDNS and NHANES) provides a detailed and insightful analysis of misreporting, including stratification by age, sex, and BMI. This is important because it allows for a nuanced understanding of how misreporting varies across different demographic groups and highlights the need for targeted interventions.
Section: Results
Comprehensive Discussion of Findings (written-content)
The Discussion section provides a thorough and comprehensive discussion of the study's findings, effectively placing them within the context of existing literature and addressing potential implications. This is important because it helps to establish the study's contribution to the field and highlights the significance of the findings for future research and practice.
Section: Discussion
Provide Context on Misreporting Implications (written-content)
The Abstract lacks sufficient context on the implications of misreporting for research and public health. Adding this context would significantly improve the study's impact by emphasizing the practical value of the findings and the importance of accurate dietary data.
Section: Abstract
Expand on Machine Learning Model Comparisons (written-content)
The Results section lacks sufficient detail on the comparison between classical regression and machine learning models. Providing a more comprehensive comparison would improve the study's methodological rigor by ensuring transparency in the model selection process and providing a more complete picture of the analytical approach.
Section: Results
Expand on the Implications of Findings for Dietary Guidelines (written-content)
The Discussion section lacks a thorough exploration of the implications of the findings for dietary guidelines. Adding this discussion would significantly improve the study's impact by emphasizing the need for evidence-based recommendations that account for the limitations of self-reported data.
Section: Discussion
Lack of Legend in Figure 1 (graphical-figure)
Figure 1 lacks a legend to differentiate between datasets and age groups. Adding a legend would significantly improve the clarity and interpretability of the graphs, making it easier for readers to understand the presented data.
Section: Results
Lack of Trend Line Description in Figure 2 (graphical-figure)
Figure 2's caption does not mention that the red lines are trend lines, nor does it specify the method used to generate them. Including this information would improve the clarity and scientific validity of the figure, as the type of trend line used can impact the visual interpretation of the data.
Section: Results
Clarity of Axes Labels in Figure 3 (graphical-figure)
While the axes labels in Figure 3 are generally clear, the x-axis label could be improved by specifying that it represents the 'Percentage of total energy from macronutrient'. This would enhance consistency with previous tables and improve the overall clarity of the figure.
Section: Discussion

Section Analysis

Abstract

Key Aspects

Strengths

Suggestions for Improvement

Introduction

Key Aspects

Strengths

Suggestions for Improvement

Results

Key Aspects

Strengths

Suggestions for Improvement

Non-Text Elements

Table 1 | Significant terms in the general linear model analysis (10 decimal...
Full Caption

Table 1 | Significant terms in the general linear model analysis (10 decimal places) predicting TEE

Figure/Table Image (Page 3)
Table 1 | Significant terms in the general linear model analysis (10 decimal places) predicting TEE
First Reference in Text
The derived significant predictors and their regression coeffi-cients are reported in Table 1.
Description
  • Purpose of the table: This table shows the results of a statistical analysis called 'general linear model analysis'. You can think of this like drawing a line through a cloud of points on a graph to see if there's a relationship between different things. Here, they're looking at what factors might be related to something called 'Total Energy Expenditure' (TEE), which is the amount of energy (like calories) a person uses in a day.
  • Content of the table: The table lists different factors that the analysis found to be important in predicting TEE. For each factor, it gives a 'coefficient,' which is like the slope of the line in our graph analogy – it tells us how much TEE is expected to change if that factor changes. It also shows the 'standard error' of the coefficient, which is a measure of how precise that estimate is. Additionally, it includes a 'T-value,' which helps determine if the factor is statistically significant, and a 'P-value,' which is the probability of seeing the observed relationship (or a stronger one) if there was actually no real relationship between the factor and TEE. A low P-value (typically less than 0.05) suggests that the relationship is statistically significant, meaning it's unlikely to have occurred by chance.
  • Specific factors listed: The factors listed in the table are things like body weight, height, age, sex, elevation, and ethnicity. 'In[BW (kg)]' refers to the natural logarithm of body weight in kilograms. The natural logarithm is a mathematical function that helps to transform the data in a way that makes it more suitable for this type of analysis. Similarly, 'In[Elevation (m)]' is the natural logarithm of the elevation where the measurement was taken, in meters. 'Sex' is a categorical variable, likely coded as male or female. 'Ethnicity' is also a categorical variable, with categories like 'A' for African, 'AA' for African living outside Africa, 'AS' for Asian, 'W' for White, 'H' for Hispanic, and 'NA' for not available.
  • Precision of the coefficients: The coefficients in the table are reported to 10 decimal places. This level of precision is unusual in many scientific fields, but the authors state later in the paper that they did this for reproducibility, so that others can get the exact same results if they use their model.
Scientific Validity
  • Model Specification: The authors have included a comprehensive set of predictors in their model, including demographic, anthropometric, and environmental variables. The inclusion of interaction terms (e.g., Height x In[Elevation (m)]) is appropriate and suggests a thorough exploration of potential relationships. However, the rationale for including specific interaction terms could be more explicitly stated.
  • Statistical Significance: The reported T-values and P-values allow for a clear assessment of the statistical significance of each predictor. The use of a general linear model is appropriate given the nature of the dependent variable (TEE) and the mix of continuous and categorical predictors.
  • Precision of Coefficients: While reporting coefficients to 10 decimal places is unconventional, the authors justify this decision based on the need for precise replication of their predictive model. This level of precision, although unusual, does not inherently invalidate the scientific rigor of the analysis. It is crucial, however, that the authors provide a sensitivity analysis demonstrating the impact of this precision on the model's predictions.
Communication
  • Clarity of Column Headers: The column headers are generally clear and provide sufficient information to understand the table's contents. However, 'SE coefficient' could be more explicitly labeled as 'Standard Error of Coefficient' for better clarity.
  • Footnotes: The footnote defining the ethnicity abbreviations is helpful. However, providing a brief explanation of the coding for 'Sex' within the table or in a footnote would further enhance clarity.
  • Caption Clarity: The caption is concise but could be slightly more descriptive. For example, it could be revised to: 'Table 1 | Significant terms and their coefficients from the general linear model analysis predicting Total Energy Expenditure (TEE), presented with 10 decimal places for reproducibility.'
Table 2 | Summary of observations inside and outside the tolerance limits in...
Full Caption

Table 2 | Summary of observations inside and outside the tolerance limits in the NDNS and NHANES datasets

Figure/Table Image (Page 4)
Table 2 | Summary of observations inside and outside the tolerance limits in the NDNS and NHANES datasets
First Reference in Text
Using the predictive equations developed above, the number and percentage of individuals that fell outside the predicted limits (both over and under) and within the predicted limits are shown in Table 2, stratified by data source, age (adults versus children) and sex.
Description
  • Purpose of the table: This table summarizes how well the predictions from their equation match up with real-world data from two large dietary surveys, the NDNS (National Diet and Nutrition Survey) and NHANES (National Health and Nutrition Examination Survey). It's like checking if a weatherman's forecast (the equation's prediction) accurately reflects the actual weather (the survey data).
  • Tolerance Limits: The 'tolerance limits' refer to the range of values within which the researchers expect the real-world data to fall, based on their equation. It's like setting a margin of error around the prediction. If the weatherman predicts a temperature of 20 degrees Celsius, he might say the actual temperature will likely be between 18 and 22 degrees. That range is the tolerance limit.
  • Organization of the table: The table is organized by dividing the survey participants into groups based on which survey they were part of (NDNS or NHANES), whether they were adults or children, and their sex (male or female). For each group, the table shows how many people had reported dietary intakes that fell within the predicted range (inside the tolerance limits), below the predicted range (underestimated), and above the predicted range (overestimated). These counts are also shown as percentages of the total number of people in each group.
  • Interpretation of the data: If the equation is a good predictor, most people's reported intakes should fall within the tolerance limits. If a large percentage of people fall outside the limits, it suggests that the equation may not be accurately reflecting real-world dietary intake, or that there's a lot of misreporting in the surveys. For example a high percentage of 'underestimated' would suggest that many people are reporting eating less than what the equation predicts they should need based on factors like their weight, height, and age.
Scientific Validity
  • Appropriateness of Tolerance Limits: The scientific validity of this table hinges on the appropriateness of the tolerance limits used. The authors have previously described how these limits were derived (95% prediction intervals), which is a statistically sound approach. However, the validity of applying these limits to assess misreporting relies on the assumption that deviations from the predicted TEE primarily reflect misreporting rather than individual variability or other factors not captured by the model.
  • Stratification by Data Source, Age, and Sex: Stratifying the results by data source, age, and sex is crucial for identifying potential biases or differences in the performance of the predictive equation across different populations. This allows for a more nuanced interpretation of the results and helps to pinpoint specific groups where misreporting may be more prevalent. The choice of these stratification variables is justified given their known associations with dietary intake and reporting behaviors.
  • Use of Number and Percentage: Presenting both the number and percentage of individuals within each category is helpful for interpretation. Percentages provide a standardized way to compare across groups of different sizes, while raw numbers give a sense of the actual sample sizes involved.
Communication
  • Clarity of Column Headers: The column headers are relatively clear but could be improved. 'Number underestimated' and 'Number overestimated' could be more explicitly defined as 'Number below the lower tolerance limit' and 'Number above the upper tolerance limit,' respectively. Similarly, 'Number within range' could be clarified as 'Number within tolerance limits.'
  • Caption Clarity: The caption is generally clear and informative. It could be slightly improved by explicitly stating that the tolerance limits are based on the 95% prediction intervals of the predictive equation.
  • Footnote: The footnote is helpful in explaining what the table shows. However, it could be made more informative by briefly mentioning the years the datasets cover, which is relevant context for interpreting the results.
  • Overall Readability: The table is well-organized and relatively easy to read. The use of bold font for the main categories (e.g., Male children, Female children) enhances readability.
Fig. 1 | Misreporting in relation to age, BMI and sex. a, Comparison of the...
Full Caption

Fig. 1 | Misreporting in relation to age, BMI and sex. a, Comparison of the difference between predicted TEE and self-reported energy intake (EI) in the NDNS (n = 12,694) and NHANES (n = 5,873) datasets in relation to age for children (≤16 yr) and adults (>16 yr). b, Comparison of the difference between predicted TEE and self-reported energy intake in the same datasets in relation to BMI for children (≤16 yr) and adults (>16 yr). Negative values show observations lower than prediction and positive values show prediction higher than observation.

Figure/Table Image (Page 5)
Fig. 1 | Misreporting in relation to age, BMI and sex. a, Comparison of the difference between predicted TEE and self-reported energy intake (EI) in the NDNS (n = 12,694) and NHANES (n = 5,873) datasets in relation to age for children (≤16 yr) and adults (>16 yr). b, Comparison of the difference between predicted TEE and self-reported energy intake in the same datasets in relation to BMI for children (≤16 yr) and adults (>16 yr). Negative values show observations lower than prediction and positive values show prediction higher than observation.
First Reference in Text
We plotted the difference between the survey estimate of daily energy intake and the predicted TEE as a function of age and body mass index (BMI) for both the NDNS and NHANES datasets (Fig. 1).
Description
  • Overall Purpose: This figure is trying to show how well people's self-reported food intake matches up with what a scientific equation predicts they should be eating. The equation predicts Total Energy Expenditure (TEE), which is the number of calories a person burns in a day. The researchers are comparing this prediction to self-reported energy intake (EI), which is what people say they eat in dietary surveys. The difference between these two (predicted TEE - self-reported EI) is an indication of potential 'misreporting' - either under-reporting (eating less than they say) or over-reporting (eating more than they say).
  • Structure of the Figure: The figure is divided into two parts, labeled 'a' and 'b'. Each part contains four graphs. Part 'a' looks at the relationship between misreporting and age, while part 'b' looks at the relationship between misreporting and Body Mass Index (BMI), which is a measure of body fat based on height and weight. Each graph is a scatter plot, with each dot representing a person in the study. The graphs are further split by the dataset used (NDNS or NHANES) and whether the participants were children or adults.
  • X and Y Axes: In part 'a', the x-axis (horizontal) represents age in years, while in part 'b', it represents BMI. In both parts, the y-axis (vertical) represents the difference between predicted TEE and self-reported EI. A value of 0 on the y-axis means that the predicted TEE and self-reported EI are the same. Negative values mean that people are reporting eating less than the equation predicts (under-reporting), while positive values mean they are reporting eating more (over-reporting).
  • Interpretation of the Data Points: Each dot on the scatter plots shows an individual's data. For example, in part 'a', a dot on the NDNS-Adults graph with an x-axis value of 40 and a y-axis value of -5 would represent a 40-year-old adult in the NDNS study who reported eating 5 megajoules (a unit of energy) less per day than the equation predicted. The red line on each graph is a trend line, which is like drawing a line through the middle of the dots to see the general pattern. If the trend line slopes downwards, it means that as age or BMI increases, the difference between predicted TEE and self-reported EI tends to become more negative (more under-reporting).
Scientific Validity
  • Appropriateness of Visualization: Using scatter plots with trend lines is an appropriate way to visualize the relationship between continuous variables like age, BMI, and the difference between predicted and reported energy intake. This allows for a visual assessment of the magnitude and direction of misreporting across different age and BMI groups.
  • Statistical Analysis: The reference text indicates that the authors plotted the difference between predicted and reported energy intake, but it doesn't specify the method used to generate the trend lines. It's crucial to know whether these are simple linear regressions or if a more sophisticated smoothing technique was employed. The choice of method can influence the interpretation of the trends.
  • Sample Size: The large sample sizes from the NDNS and NHANES datasets provide robust data for this analysis. However, it's important to consider potential biases or limitations inherent in these datasets, such as the reliance on self-reported dietary intake.
Communication
  • Clarity of Axes Labels: The axes labels are generally clear and informative. The y-axis label could be slightly improved by specifying the units (MJ d-1) for the difference between predicted TEE and self-reported EI.
  • Legend: The figure lacks a legend to differentiate between the NDNS and NHANES datasets and between children and adults. Adding a legend would significantly improve the clarity and interpretability of the graphs.
  • Caption: The caption is relatively clear but could be more concise. It could also benefit from explicitly stating that the red lines represent trend lines.
  • Visual Clutter: The use of different colors and symbols for each dataset and age group, combined with the trend lines, creates some visual clutter. Using a more minimalist color scheme or separating the graphs for children and adults could improve readability.
  • Trend Line Description: The caption should mention that the red lines are trend lines, and the method used to generate them should be specified in the methods section.
Table 3 | Relationships between the discrepancy of intake to expenditure and...
Full Caption

Table 3 | Relationships between the discrepancy of intake to expenditure and self-reported dietary macronutrient composition

Figure/Table Image (Page 5)
Table 3 | Relationships between the discrepancy of intake to expenditure and self-reported dietary macronutrient composition
First Reference in Text
Next, we explored the relationship between the discrepancy in energy intake and the proportional macronutrient composition (percentage energy) of the reported diet (Table 3).
Description
  • Purpose of the Table: This table explores whether the tendency for people to over- or under-report their food intake is related to the types of food they eat. Specifically, it looks at whether the difference between what people report eating and what an equation predicts they need (the 'discrepancy') is linked to the proportion of their diet that comes from carbohydrates, protein, and fat. These three are called 'macronutrients' and are the main components of food that provide energy.
  • Structure of the Table: The table is divided into four sections, each representing a different set of data: the full NDNS dataset, the screened NDNS dataset, the full NHANES dataset, and the screened NHANES dataset. 'Screened' here likely refers to removing data points that were considered unreliable based on some criteria, like falling outside the tolerance limits mentioned in previous tables. Each section shows the results of a statistical analysis called 'multiple regression analysis'. This is a method used to examine the relationship between a dependent variable (in this case, the discrepancy between reported and predicted energy intake) and several independent variables (the percentage of energy from carbohydrates, protein, and fat).
  • Key Terms Explained: 'Coefficient' in this context refers to the estimated change in the discrepancy (in kilojoules per day) for a one-unit change in the percentage of energy from each macronutrient. For example, a coefficient of -207.3 for 'Percentage protein' in the full NDNS dataset means that for every 1% increase in the proportion of protein in the diet, the discrepancy between reported and predicted intake is estimated to decrease by 207.3 kJ/day (meaning more under-reporting). 'SE coefficient' stands for the standard error of the coefficient, which is a measure of the precision of the estimate. 'P-value' is the probability of observing the relationship (or a stronger one) if there was actually no real relationship between the macronutrient and the discrepancy. A low P-value (typically less than 0.05) suggests that the relationship is statistically significant. 'R²' is a measure of how well the model fits the data, with higher values indicating a better fit.
Scientific Validity
  • Appropriateness of Statistical Method: Multiple regression analysis is an appropriate method for examining the relationship between the discrepancy in energy intake and the proportional macronutrient composition of the diet. The use of both full and screened datasets allows for an assessment of the robustness of the findings.
  • Interpretation of Coefficients: The interpretation of the coefficients is crucial. The authors should provide a more detailed discussion of the implications of the positive and negative coefficients for different macronutrients. For instance, they should discuss potential reasons why a higher reported protein intake is associated with a greater discrepancy (more under-reporting).
  • Consideration of Confounding Factors: While the table presents the results of multiple regression, which adjusts for the other included macronutrients, there may be other confounding factors that influence both macronutrient composition and the discrepancy in energy intake. These potential confounders should be acknowledged and discussed.
  • Comparison of Full and Screened Datasets: Comparing the results from the full and screened datasets is valuable for assessing the impact of removing potentially unreliable data points. The authors should provide a more detailed comparison of these results and discuss any notable differences.
Communication
  • Clarity of Column Headers: The column headers are generally clear and informative. However, 'Term' could be more explicitly labeled as 'Predictor Variable' or 'Macronutrient.'
  • Caption Clarity: The caption is concise but could be more descriptive. It could be revised to: 'Table 3 | Results of multiple regression analyses examining the relationships between the discrepancy of reported energy intake to predicted expenditure and the self-reported dietary macronutrient composition (percentage of total energy) in the NDNS and NHANES datasets, using both full and screened data.'
  • Table Organization: The organization of the table into four sections is logical and facilitates comparison across datasets and data treatments (full vs. screened). However, the table is quite dense, and the use of bold font or shading could help to visually separate the different sections.
  • Explanation of Screening: The table would benefit from a brief explanation of the screening criteria used to define the 'screened' datasets. This information could be included in a footnote or in the methods section.
  • Units: It would be helpful to include the units for the coefficients (kJ/day per 1% change in macronutrient) in the column header or a footnote.
Fig. 2 | Misreporting and macronutrient intake. a-c, The discrepancy between...
Full Caption

Fig. 2 | Misreporting and macronutrient intake. a-c, The discrepancy between the predicted TEE and the reported energy intake in the NHANES and NDNS surveys plotted against the self-reported intakes of fat (a), protein (b) and carbohydrates (c) as a percentage of the total energy. For each macronutrient, the top two plots show data from the whole sample (full data) and the bottom two plots show the data from the sample screened to include only those individuals within the predictive interval of the equation (screened). Significant effects in the whole sample were severely attenuated in the screened sample (see Table 3 for regression details).

Figure/Table Image (Page 6)