Reproducibility of Dietary Intakes of Macronutrients, Food Groups, and Dietary Patterns in 211,050 Adults from the UK Biobank

Why Reproducibility Matters

Reproducibility (or reliability) quantifies the consistency of a measurement across time. In nutritional epidemiology, low reproducibility dilutes true dietdisease associations, leading to regressiontothemean bias** and attenuation of risk estimates. Understanding the magnitude of withinperson variation for macronutrients, food groups, and dietary patterns helps researchers:

• Choose appropriate exposure metrics (e.g., averages of repeated measures).
• Calculate correction factors for measurement error.
• Interpret null findings more accurately.

Methods Overview

Participants completed the Oxford WebQ up to three times, with a median interval of 3.2years between consecutive recalls. Energy and nutrient intakes were estimated using the McCance &White food composition tables. Food groups (e.g., fruits, vegetables, red meat, processed meat, dairy) were derived by aggregating individual items. Two dietary patterns were examined:

1. Principal Componentderived Western pattern (high in red/processed meat, refined grains, sugary drinks).
2. Principal Componentderived Prudent pattern (high in fruits, vegetables, whole grains, fish).

Reproducibility was assessed with:

• Intraclass correlation coefficients (ICCs) for continuous variables.
• Weighted kappa for categorical foodgroup frequencies.
• Spearman correlation for pattern scores.

All estimates were age and sexadjusted and expressed with 95% confidence intervals (CIs).

Macronutrients

Table 1 summarises the ICCs for total energy and the main macronutrients.

Energy and protein showed the highest reproducibility, whereas alcohol intake was the least stable. These ICCs are comparable with previous repeatdiet studies that used food frequency questionnaires (FFQs) but are modestly lower than those observed for biomarkers (e.g., urinary nitrogen for protein).

FoodGroup Intakes

Weighted kappa values for the most frequently consumed food groups are shown in Table2. Categories were defined as low, moderate, and high based on tertiles of intake.

Red and processed meats displayed the strongest agreement, while beverage consumption was the most variable. The modest kappas indicate that a single 24hour recall cannot reliably rank individuals for most food groups; averaging two or more recalls improves classification markedly.

Dietary Patterns

Pattern scores were derived from principal component analysis (PCA) applied to the 24hour recall data. The first component represented a Western pattern, the second a Prudent pattern. Spearman correlation coefficients between the first and second assessments are displayed in Table3.

Correlations around 0.5 suggest moderate stability. When pattern scores were averaged across all available repeats (up to three), the reliability increased to 0.65, supporting the use of cumulative averages in prospective analyses.

Impact of Repeated Measures

To illustrate the benefit of multiple recalls, ICCs were recalculated using the mean of two and three assessments (Figure1). For total energy, the ICC rose from 0.47 (single recall) to 0.61 (two recalls) and 0.68 (three recalls). Similar gains were observed for most nutrients and food groups.

Figure1. Intraclass correlation coefficients for total energy intake as a function of the number of WebQ repetitions.

These findings reinforce the recommendation to use the average of at least two dietary assessments when investigating longterm health outcomes.

Strengths and Limitations

• Strengths: Very large sample, nationwide coverage, webbased automated recalls reducing interviewer bias, and availability of up to three repeats per participant.
• Limitations: 24hour recalls capture shortterm intake and may miss episodic foods; participants who completed multiple recalls were slightly healthier and more educated, possibly limiting generalisability; residual measurement error remains even after averaging.

Implications for Future Research

Researchers using UK Biobank dietary data should consider the following:

1. Use the mean of all available WebQ recalls to improve exposure reliability.
2. Apply regressioncalibration or simulationextrapolation (SIMEX) methods to correct for remaining measurement error.
3. When studying foods with low repeatability (e.g., sugary drinks), combine WebQ data with objective biomarkers where possible.
4. Report ICCs or reliability metrics alongside effect estimates to aid interpretation.

By accounting for withinperson variability, the UK Biobank can provide more precise estimates of the relationship between diet and chronic diseases such as cardiovascular disease, diabetes, and cancer.

Key TakeHome Messages

• Macronutrient intakes show moderate reproducibility (ICCs0.450.50); protein and total energy are the most stable.
• Foodgroup classification is less reliable, especially for beverages and fruits.
• Dietary pattern scores are moderately reproducible (Spearman r0.5); reliability improves with repeated measures.
• Using the average of two or three 24hour recalls substantially raises reliability, reducing attenuation bias in epidemiologic analyses.