Data

This study uses data from the UK Biobank. UK Biobank is a population-based longitudinal study that recruited roughly 500,000 participants aged between 40 and 69 when they joined UK Biobank (from 2006 to 2010). It follows the health and wellbeing of the participants over a 17-year period with future data releases scheduled. The volunteer participants of UK Biobank completed a full baseline assessment, including self-reported measurements via touch-screen questionnaires as well as a verbal interview collecting a wide range of information on socio-demographic factors, lifestyle, and behaviours (i.e., history of smoking and sleep duration), and medical history. Physical measurements (i.e., height, weight, spirometry, blood pressure, heel bone density), blood and urine samples were also taken.

UK Biobank protocols and study details can be found on the UK Biobank website (https://www.ukbiobank.ac.uk/). UK Biobank is not representative of the general population with evidence of a ‘healthy volunteer’ selection bias, details of which are available online on the UK Biobank website (http://www.ukbiobank.ac.uk/wp-content/uploads/2011/11/UK-Biobank-Protocol.pdf).

We take advantage of a longitudinal dataset conducted in the UK by using repeated measures on an individual to control for time invariant unobserved heterogeneity and therefore reduce the bias often associated with studies that use a cross-sectional design. Whilst results are generated for all food types and nutrients, the paper presents selected food groups for illustrative purposes. Additional nutrient types are presented in the supplementary materials.

Dietary assessment

Data on dietary intake were collected by all UK Biobank participants who provided a valid email address at recruitment. Participants were invited to complete the 24-hour online dietary assessment (Oxford WebQ), which is a web-based 24-h dietary assessment tool developed and evaluated for use in large population studies (www.ceu.ox.ac.uk/research/oxford-webq). The Oxford WebQ was collected toward the end of the baseline recruitment period of UK Biobank (2009–2010). Follow ups were conducted on up to four separate occasions (February 2011 to April 2011; cycle 2: June 2011 to September 2011; cycle 3: October 2011 to December 2011; cycle 4: April 2012 to June 2012) [15]. For the purpose of this study, we focus on respondents in the UK Biobank who have completed at least one Oxford WebQ. Our sample size is therefore 185,611 individuals.

The Oxford WebQ presents participants with 21 broad food groups, expanding to offer 206 commonly consumed food and 32 types of drinks. The participants are prompted to select the number of portions consumed over the previous 24 h, mostly from predefined categories offered to them (www.ceu.ox.ac.uk/research/oxford-webq). Until recently, the food composition table (FCT) and portion size used for the Oxford WebQ has been the UK McCance and Widdowson’s “The Composition of Foods 6th edition (2002). This has now been replaced by the UK Nutrient Databank (UKNDB) (2013), which provides food composition data measured closer in time to when participants completed the questionnaire in UK Biobank (2009–2012).

In the Oxford WebQ, nutrients are automatically estimated via built-in algorithms and food composition data [16]. For the purpose of this study, we focus on total energy intake, total fat, saturated fatty acids (SFA), monounsaturated fatty acids (MUFA), polyunsaturated fatty acids (PUFA), carbohydrates, total sugars, and fibre. The Oxford WebQ has been validated against biomarkers [17] and compared to interviewer-administered 24 h recalls [18] and showed acceptable reproducibility when using at least 2 dietary assessments [19, 20].

Study design

Following the previous literature, we removed participants with implausible energy intakes [21]. These were defined as (men:  17,573 kJ/days or  4200 kcal/days); women:  14,644 kJ/days or  3500 kcal/days). Recorded food and drinks from the Oxford WebQ were classified into 44 groups according to their nutrient profile and the classification used in the UK National Diet and Nutrition Survey (NDNS). This allows us to make comparisons between the change in nutrient intake and the average nutrient intake of that food group in the NDNS. For the purpose of this study, we use data from Wave 11 of the NDNS 2018/19.

As mentioned previously, reductions in consumption of one food group may be correlated with reductions or increases in consumption of other food groups. To estimate the effect of a reduction in one food group on overall nutrient intake we use a fixed effects panel model for each food group and nutrient separately. That is, formally:

where ({I}_{int}) is intake by individual i of nutrient n at time period t, ({C}_{ift}) is consumption by individual i of food group f at time t. ({delta }_{i}) is a fixed effect for individual i which captures individual heterogeneity and ({delta }_{t}) is a time fixed effect which captures changes over time that are applicable to everyone. ({epsilon }_{ift}) is a random error term. The (beta) coefficient is thus the estimate of the effect of a change in consumption of food group f on intake of nutrient n. The fixed effects model has particular strengths over a random effects model because it removes any bias from food preferences. For example, it may be that people who eat more chocolate also just have a taste for sugary foods generally. A model which didn’t control for individuals’ preferences would see a positive relationship between the number of chocolate bars eaten and sugar consumption, whereas at the individual level people may also be less likely to drink a sugary drink when they eat a chocolate bar and so the relationship is not as strong.

Utilising data from wave 11 of the NDNS food level diary data, we calculate the average nutritional content of foods within each of the 44 food groups. This allows us to compare our estimated change in nutrient intake against the average nutritional value of that particular food group consumed in wave 11.