Day-Time Patterns of Carbohydrate Intake in Adults by Non-Parametric Multi-Level Latent Class Analysis—Results from the UK National Diet and Nutrition Survey (2008/09–2015/16)

The UK National Diet and Nutrition Survey Rolling Program (NDNS RP 2008/09–15/16) [12] is designed as a nationally representative sample to assess the diet, nutrient intake and nutritional status of the general population living in the UK. Details of the rationale, design and methods of the survey have been described elsewhere [13,14]. In brief, for years 1 to 8 combined, a sample of 39300 addresses was randomly selected from 1438 postcode sectors or Primary Sampling Units (PSUs). Response rates for completion of the required diet diary (three to four consecutive days) were 56%, 53% and 53% for years 1 to 4, 5 to 6 and 7 to 8, respectively. A total of 6155 adults (2537 men and 3618 women, age ≥ 19-year-old) were retained in the current study. Ethical approval for the NDNS RP was obtained from the Oxfordshire Research Ethics Committee [13].

Participants were asked to keep a record of everything eaten or drunk over four consecutive days. Accordingly, there were 24,483 observation days collected in the NDNS RP dataset. In the diet diary, participants were also asked to write down the portion sizes in household measures. Detailed diary checking was performed by trained staff to code and convert the food consumption into energy and nutrients intake using the DINO (Diet in Nutrients Out) [15] system. Intakes of nutrients were calculated from the food consumption records using a specially adapted Nutrient Databank [16] originally developed for the Dietary and Nutritional Survey of British Adults [17]. Further details of data coding and editing are outlined in Appendix A of the NDNS RP official reports [13,14].

The consumption of CH was defined as total sugar plus starch intake (gram), fiber was not included as it was deemed not to contribute any energy intake in the NDNS RP data [18]; time across a typical survey day was classified according to the standard NDNS classification which consists of 7 pre-defined time slots: 6–9 am, 9–12, 12–2 pm, 2–5 pm, 5–8 pm, 8–10 pm, 10 pm–6 am; CH consumption in each time-slot was further categorized as: no energy intake occurred; CH contributed <50% or ≥50% of energy intake within that time slot, where these thresholds reflect the current UK recommendation for daily CH intake [19].

Latent class analysis (LCA) is a statistical technique that performs unsupervised classifications based on observed categorical variables [20]. Multilevel latent class analysis (MLCA) [21,22,23] is an extension of LCA which allows for the repeated measurement of categorical variables, here represented by CH intake over 3–4 days nested within individuals. MLCA extracts day-level latent classes which are then used as indicators for a second latent class model at the individual level. Therefore, under MLCA models, individuals are allowed to have different probabilities of following alternative DEPs during their survey period. In particular, we employed non-parametric MLCA approach which relaxed the assumption that the multi-level model random intercepts follow a normal distribution and is less computationally demanding [21,24].

The following sequential modelling strategy was applied [22]: First, a series of traditional LCA models were fitted to determine the number of classes at the day-level, ignoring the hierarchical structure of the observations; next, a series of MLCA models were fitted to derive the individual-level classes accounting for such hierarchical structure; last, sensitivity analyses were conducted to investigate the effect of changing day level classification on the individual level classification. For both LCA and MLCA, the class assignment was determined by the maximum posterior probability given the model using the Bayes theorem.

The number of classes at day level was selected by the following rules: 1) The evaluation of model fit indices, including the Bayesian information criterion (BIC) and entropy; 2) the Lo-Mendell-Rubin Likelihood Ratio Test (LMR-LRT) [25,26] which compares q vs. q-1 classes models, where q is the number of latent classes; and most importantly; 3) clear class separation as well as pattern interpretability. LCA and MLCA models were fitted in MPlus (Version 8,3, Muthén & Muthén: Los Angeles, CA, USA) [27]. Example Mplus codes for fitting the MLCA model are given in Appendix A.

Background information of the participants on socio-demographic variables were collected (including smoking habits, ethnicity, education level, living with a partner or not, etc.). Participants also had their height, weight and waist circumferences (WC) measured during the interview visit. Recent Physical Activity Questionnaire (RPAQ) [28] was used to estimate physical activity. Blood samples were taken from the participants to measure their fasting blood glucose and HbA1c. T2D cases were defined if any of the following criteria were met: 1) fasting blood glucose ≥7 mmol/L; 2) HbA1c ≥6.5%; 3) self-report of being diagnosed with T2D. We also identified undiagnosed T2D participants if they met either of the first two criteria but did not report having T2D.

Individual-level point estimates of characteristics and 95% confidence intervals (CIs) were estimated by applying the relevant sampling weights [15,16]. Descriptive statistics of participants were presented as weighted means (95% CI) or weighted percentages (95% CI). Fasting blood glucose, HbA1c and average physical activity duration (hours-per-day) were log-transformed to improve normality. Weighted estimates of energy consumption (and their composition) across the 7 day-time slots were calculated for each individual-level latent class.

Associations between individual-level CH classes and T2D were explored using survey-adjusted multivariable logistic regression models. Potential confounders were selected as covariates if they had an association with both individual-level CH classes and T2D. Consequently, a crude model; a model adjusted for age (continuous), sex (men, women), BMI (continuous), smoking status (past, current, never), total energy intake (continuous), alcohol intake (continuous); and a model adjusted for the same covariates except with BMI being replaced with WC were fitted. We also conducted the same analyses taking (i) only the self-reported T2D status as the outcome among the total sample (independently of their fasting glucose and HbA1c), or (ii) undiagnosed T2D (those diagnosed by either HbA1c or fasting glucose but did not report diabetes) to see whether diagnosis was associated with reported food choices in survey members. Except for MLCA, all the statistical analyses were performed in Stata (Version 15.1, StataCorp LLC, College Station, TX, USA) [29].

The majority of the participants (n = 6031, 97.9%) succeeded to provide a 4-day diary for the current analyses. More women (n = 78, 2.2%) than men (n = 46, 1.8%) discontinued their diary on day 3 or skipped 1 day of their diet diary. After examining and comparing details of different solutions (discussion of how the results were chosen, and figures from other competing solutions are presented in Appendix B), three CH eating day patterns emerged from the collected diet diaries (Figure 1).

Class 1 (Figure 1A) was defined as “high percentage CH day” since the probabilities (p) of CH contributing ≥50% of energy was relatively high across day time slots. In class 1 days, the probabilities were greater than 0.6 of having more than 50% of calories from CH between 6 a.m. to 9 a.m. (breakfast), 9 a.m. to 12 noon (morning snack) and 2 p.m. to 5 p.m. (afternoon snack). Moreover, during late evening and night time, the probabilities of having CH rich food were still relatively high (0.412 and 0.246, respectively).

Class 2 days (Figure 1B) were named as “low percentage CH day” because between 9 a.m. and 10 pm, the probability of carbohydrates contributing <50% of energy was always the highest. In class 2 days, participants also tended to have morning snacks (p = 0.079 of no energy intake), although these may also be interpreted as having a long or late breakfast in these mornings. Also, the probability of no intake at night was the lowest for Class 2 days (p = 0.59).

Class 3 days (Figure 1C) were defined as “regular meal CH day” because: 1) The probability of no energy intake was virtually null at lunch (p = 0.019) and dinner time (p = 0.034); 2) the probabilities of no energy intake between 9 a.m. and 12 noon and between 2 p.m. and 5 p.m. were also the highest (p = 0.401 and p = 0.659, respectively).

Detailed characteristics of the three CH eating day patterns can be found in. In brief, consumption of total energy, CH, sugar, starch and non-milk extrinsic sugar was the highest among Class 1 days. On the other hand, intake of protein, fat, and alcohol were the highest in the Class 3 days. Therefore, we anticipated that if one’s diet diary contains higher proportions of Class 1 days, or lower proportions of Class 3 days he/she would probably consume more CH in general. Supplementary Tables S1 and S2

Figure 2 showed the results from the chosen MLCA model where low-CH eaters (LCE, 28.1%), moderate-CH eaters (MCE, 28.8%) and high-CH eaters class (HCE, 43.1%) were defined. In the LCE group, 62.3% of days were classified as Class 3, and only 20.5% as Class 1. MCE had comparable proportions of day Class 1 (42%) and 3 (40%) and much lower one of Class 2 (18.0%). Among HCE, nearly 50% of their days were classified as Class 1 and approximately 25.8% of days of their dietary diary were of Class 3 and 24.1% were of Class 2. Individual level classification reflects differences in the probability of following a specific day-level class so that individuals with similar probabilities of following the day level classes were grouped together. During their 4-day-diary, all types of CH eaters have more than one class of CH eating day, and this is reflected in sizeable conditional probabilities (proportions) for each type of day.

Estimation of the population average nutrient intake is listed in Table 1. Specifically, the detailed consumption of CH over the time of day in the adult population and according to different types of CH eaters is estimated accounting for complex survey design. Overall, the targeted population in the survey consumed 7668.9 kJ on average, the contribution of their total energy from protein, fat, CH, and alcohol were 16.6%, 33.1%, 45.8%, and 4.4%, respectively. Total energy contributed by CH was close to 50% among HCE but was only 40.6% among LCE. In terms of components of CH consumed at each time slot, HCE consumed more than twice as LCE and nearly 4 times as MCE amount of sugar (37.9 g, 95% CI: 36.8, 39.2) and non-milk extrinsic sugar (i.e., free sugar, 11.1 g 95%CI: 10.7, 11.6) between 6–9 a.m. MCE consumed more sugar and starch outside the traditional meal times: between 9–12 noon, 2–5 p.m., 8–10 p.m. and 10 p.m.–6 a.m. The mean total fiber consumption was the highest for HCE (14.7 g, 95% CI: 14.4, 14.9) compared to MCE (12.5 g, 95%CI: 12.1, 12.9) and LCE (13.7 g, 95% CI: 13.4, 14.0).

Weighted composition in total energy intake across the time of the day for three types of CH eaters are illustrated in Figure 3. LCE had the highest total energy intake (7985.8 kJ, 95%CI: 7283.3, 8146.3) while MCE had the lowest energy intake (7341.8 kJ, 95%CI: 7172.5, 7511.2, Table 1). Moreover, in LCE (Figure 3A), CH never contribute more than 50% of their total energy over the day and mostly had the highest energy from fat out of the 3 classes of eaters. Energy contributed by alcohol in LCE was the highest across time slots and considerably higher after 8pm, namely 20.0% (8–10 pm) and 31.2% (10–6 am).

On the other hand, the energy consumption in the MCE (Figure 3B) seemed to be shifted towards later in the day compared with HCE (Figure 3C). MCE tended to have later meals and they also had the highest total energy consumption (599.7 kJ) at night (10 pm–6 am) across all three types of eaters. HCE consumed the highest total energy (929.0 kJ) between 6 a.m. to 9 a.m. in the morning and the lowest total energy between 10 p.m. to 6 a.m. (205.5 kJ). Overall, HCE consumed their energy mainly at the time slots: 6–9 am, 12–2 p.m. and 5–8 pm.

Socio-demographic characteristics of UK adults according to their individual level latent class memberships are shown in Table 2. MCEs were relatively younger. Gender distribution across the three types of CH eaters was reasonably even. The distribution of the CH eaters appears to be changing over the year of survey (p = 0.015). LCE represented 32.5% of the population in 2008 but dropped below 30% in 2016. The proportion of HCE increased from 41.2% to the highest (50.6%) in 2009, then gradually declined to 38.4% in 2016.

On average, LCE had greater BMI (27.8 kg/m²) and WC (98.9/89.9 cm in men/women) compared with 27.2, 27.3 kg/m², and 95.9/88.7, 98.1/87.2 cm respectively in MCE and HCE. MCE had the highest prevalence of being a current smoker (27.8%), shortest time of daily physical activity (geometric mean: 0.62, 95% CI: 0.57, 0.68, hours/day).

Associations of individual CH eating patterns with T2D are presented in Table 3. The prevalence of self-reported T2D was almost twice among LCE (10.6, 95%CI: 8.1, 13.8) compared to participants in HCE group. The adjusted odds of being self-reported T2D was 139% higher in LCE (OR = 2.39, 95% CI: 1.51, 3.77, p < 0.001), and 42% higher in MCE (OR = 1.42, 0.85, 2.37, p = 0.181) than in HCE. If the status of T2D considered blood tests additionally, 93 more T2D cases were identified (22, 22, and 49 undiagnosed T2D cases in LCE, MCE, and HCE, respectively). The adjusted odds of being T2D were attenuated but still 67% higher in LCE (OR = 1.67, 95% CI: 1.10, 2.52, p = 0.016), 22% higher in MCE (OR = 1.22, 95% CI: 0.77, 1.93, p = 0.407) than in HCE group. However, when the diagnosed T2D were excluded, the adjusted odds for undiagnosed T2D became 31% lower in LCE (OR = 0.69, 95% CI: 0.35, 1.35), and 15% lower in MCE (OR = 0.85, 95% CI: 0.39, 1.86) compared with HCE group.

Some statistical and epidemiological limitations in the current study merit consideration. First, we ignored the order of observation days in the MLCA models, treating the observation days as exchangeable repeated measures. However, other statistical techniques which could take the order or the longitudinal nature of the data into account, such as repeated measures latent class analysis (RMLCA) [], latent transition analysis (LTA) [], or latent class growth analysis (LCGA) [,,] are not applicable for the NDNS RP dataset as they would require that the 3 or 4 repeated measurements of diet diaries be recorded at the same time points longitudinally. ^{22 22 36 37 38}

Second, the classification of individuals to latent CH eating classes was defined by maximum posterior probability assignment rule whereby individuals are assigned to the class with the highest posterior probability of membership [], ignoring the uncertainty in the classification. However, the maximum probability rule is still able to minimize the number of incorrect assignments [,], compared to the alternative approach of multiple pseudo-class draws []. A Monte Carlo simulation study [] also found that an inclusive LCA (i.e., LCA with relevant covariates) would probably perform better than non-inclusive LCA and has the potential to reduce bias of class assignment. However, whether this advantage can be extended to MLCA in the current study is unknown. ^{39 40 41 42 41}

Third, it is not possible to differentiate between under-reporting due to ill health vs. actual misreporting in data from NDNS RP. Consequently, most previous studies using data from the NDNS RP do not adjust for under-reporting of nutrients intake in the models. Detailed data on shift work and chronotype were also not available to explore whether diurnal eating patterns differed by shift work or chronotype. Moreover, the current study only focused on characterizing patterns of CH intake, as an example of a macronutrient that may potentially influence circadian rhythms. Yet, other macronutrients may equally play a role in entraining the clock system []. Our study also did not take into account differences in CH quality in deriving the CH diurnal patterns, although we did find that LCE were consuming less “good quality” CH than HCE. Experimental studies in animal models have previously reported that rapidly digestible starch has a stronger entraining effect on peripheral clocks than slowly digestible starch []. Consequently, future studies should address how timing of intake of different types of carbohydrates may influence both short-term physiological responses and long-term health outcomes. ^{7 43}

Fourth, another important limitation is the cross-sectional study design which does not permit assessment of causal relationship between the CH eating patterns and the surveyed health outcomes. However, the sensitivity analyses we conducted on undiagnosed diabetics were indeed suggesting the presence of reverse causation (with behavior change as a result of health concerns) since HCE were in that case associated with greater prevalence of diabetes.

Finally, the findings here may not be generalizable to populations in other countries. However, datasets from other national and cultural context with a day temporal structure as well as repeated measurements in time could be explored by this approach and help us understand and summarize the complexity in human eating behavior.

Strengths of this study include the large, nationally representative sample of UK adults. The classification process through model-based, data-driven procedures minimizes reliance on researchers’ normative notions of healthy dietary patterns and can highlight multivariate/complex features that may otherwise be overlooked []. MLCA especially accounted for the multilevel structure of the data, where the 4-day diet diaries were nested within participants, capturing the day-to-day variation of the respective CH eating patterns within individuals. ⁴⁴