Engaging the complexity of diet and healthy aging in humans

Little is known about how normal variation in dietary patterns in humans affects the aging process, largely because both nutrition and the physiology of aging are highly complex and multidimensional. Here, we apply the nutritional geometry framework to data from 1560 older adults followed over four years to assess how nutrient intake patterns affect the aging process. Aging was quantified via blood biomarkers integrated to measure loss of homeostasis. Additionally, we extend nutritional geometry to 19 micronutrients. Salient results include benefits of intermediate protein and vitamin E intake. Broadly, we show that there are few simple answers of “good” or “bad” nutrients – optimal levels are generally intermediate, but dependent on other nutrients. Simpler linear/univariate analytical approaches are insufficient to capture such associations. We present an interactive tool to explore the results, and our approach presents a roadmap for future studies to explore the full complexity of the nutrition-aging landscape. Impact Statement Multidimensional nutritional analyses reveal how the association between diet and healthy aging is hard to untangle, as most nutrients have non-linear and interactive effects in humans.


56
How does what we eat affect our healthspan and longevity? The answer to this relatively 57 concise question is unavoidably complex. Conventional approaches to understanding the 58 effects of diet on health and aging, particularly in human nutrition, have usually focussed on 59 single nutrients or a handful of dietary attributes/patterns (1-3). Yet, nutrients have both 60 individual and interactive effects. For example, at the most macro-level, protein, 61 carbohydrate and fat energy sources interact to determine metabolic, physiological and 62 cognitive functioning (e.g. (4-6)). Similarly, the numerous phenotypic changes that occur with 63 age are increasingly recognized as interconnected and multidimensional (7-10). Thus 64 seemingly distinct physiological components of aging likely reflect a broader loss of 65 homeostasis in a complex dynamic system rather that independent processes (11). Such 66 interdependencies mean that the atomised interpretation of the effects of a single nutrient, 67 diet, molecular mechanism or biomarker is likely to be context-dependent (12-15); a 68 consequence being that the results of univariate studies are spurious and/or more difficult to 69 reproduce, leading to inconsistent conclusions between studies. of what can support growth and development), high-carbohydrate intake display improved 78 five systems that a previous study validated as being largely independent (10): 1) Oxygen 125 transport; 2) Liver/kidney function; 3) Leukopoiesis; 4) Micronutrients; and 5) Lipids (see 126   Table 1 for biomarkers in each score). We also calculated two other integrative, clinical-127 biomarker-based measures of biological aging: phenotypic age (PhenoAge) and  Doubal biological age (35-37). 129 130 We assessed the effects of intakes of macronutrients and micronutrients, as well as their 131 interactions, on measures of dysregulation and aging. Our primary tool was the generalised 132 additive model (GAM), a form of multiple regression that tests for non-linear 133 multidimensional effects using 'smooth' terms (38,39). We explored three-dimensional 134 effects of nutrient intake. Because GAMs can estimate non-linear effects, qualitative 135 interpretation of the sign of estimated effects comes through visualisation (rather than a 136 single regression coefficient). Here effects were visualised using the nutrient intake surfaces 137 common to the GFN. Importantly, GAMs can be used to correct for factors (e.g. 138 sociodemographic status) that might be expected to confound relationships, in the same way 139 as regression might be used in epidemiology. For each outcome we fitted a series of eight 140 models with varying levels of correction (factors explored were income, education level, age, 141 physical activity, number of comorbidities, sex and current smoking status; see text S1). 142

143
The main text contains a complete cases analysis comprising 3569 observations from 1560 144 people. In the supplementary materials (text S2 and S3) we report sensitivity analyses, where 145 we have imputed missing income data, and also analysed a more exclusive subset of the 146 complete cases dataset (see Table S1 for population summaries; the exclusive dataset 147 excluded diabetics, individuals on prescribed diets, BMI <22 or >29.9, or with substantial 148 weight fluctuation). These sensitivity analyses estimated similar effects to those in the main 149 text. However, for the exclusive dataset, in many places these effects (although qualitatively 150 similar) do not meet the criteria for statistical significance. We interpret this latter point as 151 evidence that the effects in the two datasets are similar, but that the power of the complete 152 cases analysis is required to detect statistical significance. 153 154

155
Our first model, model 1, tested for the effects of macronutrient intake (in kJ/day) on 156 outcomes without any statistical corrections (see text S1 for parameters of all models). We 157 detected statistically significant effects of macronutrient intake on liver/kidney and 158 micronutrient system dysregulation scores, as well as biological age (Figures 1a through 1c; 159 see supplementary materials Table S2 for all statistical model output). Model 1 predicted a 160 relatively minor effect of protein on liver/kidney function dysregulation, and non-linear 161 effects of all carbohydrates and fats on liver/kidney function dysregulation. Individuals who 162 consumed high (> 6000 kJ/day) or low (< 3000 kJ/day) levels of carbohydrates typically had 163 slightly elevated (around 0.25 SD above the population mean) dysregulation scores (Figure 164 1a). Very high intakes of lipids (> 4000 kJ/day; note this is within 2SD above the mean lipid 165 intake) had the highest liver/kidney function dysregulation (~0.4 SD above the mean; figure  166 1a). With regard to micronutrient dysregulation scores, individuals with moderately high 167 carbohydrate (5000 to 6000 kJ/day) and low lipid (< 2000 kJ/day) intake had minimal 168 dysregulation. Again, high lipid intake was associated with maximal dysregulation (Figure 1b). 169 Subjects with protein intake around 2000 kJ/day were predicted by model 1 to have low 170 dysregulation (Figure 1b). Biological age was predicted to be the lowest for individuals with 171 high levels of intake for all three macronutrients (Figure 1c); note that in this analysis nutrient 172 intakes are not expressed relative to any measure of individual requirements. 173

174
In model 2 we considered dysregulation score as a function of macronutrient intake relative 175 to what we define as the typical energy intake given height, weight, age, sex and physical 176 activity (text S4). We again detected effects of macronutrient intake on liver/kidney function 177 and micronutrients dysregulation and biological age score, but in addition, effects on 178 PhenoAge were detected (Figures 2a through 2d; supplementary materials Table S3). Again, 179 this analysis indicated high protein intake relative to what is typical (100% above average) 180 had low liver/kidney function dysregulation scores (Figure 2a). Interestingly, both PhenoAge 181 and biological age were predicted to be minimised at elevated carbohydrate levels and typical 182 levels of lipid and protein (Figures 1c and 1d). These effects remained after making 183 corrections for potential confounders, both excluding and including number of comorbidities 184 (models 3 and 4; supplementary materials, Tables S4 and S5), and showed a similar trend 185 within the exclusive dataset (see text S3 and Figure S3 intake on oxygen transport or lipid dysregulation scores. There were 905 models that 205 detected significant effects for at least one score, although 363 of these were solely related 206 to aging scores (i.e., no significant effect on any dysregulation score; supplementary materials 207 Table S7). There were 17 combinations for which we detected significant effects on all scores 208 (except oxygen transport and lipid dysregulation). Interestingly all 17 models contained a-209 tocopherol (vitamin E) as one of the three micronutrients. Arguably any one of these 210 micronutrient combinations may be of interest and could warrant further investigation in 211 response to a priori hypotheses (note results from all models detecting significant interactions 212 can be found at http://nuage.cohenlab.ca). Here however we focus on the effects of a-213 tocopherol, vitamin C and trans-fatty acids because this combination had the highest mean 214 percent deviance explained across all scores. 215

216
In models 5 and 6 we tested for effects of these three micronutrients with correction for 217 confounders, excluding and including comorbidities respectively. In these models we 218 detected effects of micronutrients on leukopoiesis, liver/kidney function, micronutrients and 219 global dysregulation (supplementary materials Tables S8 and S9). The effects varied slightly 220 across the four systems. (Figures 5a through 5d for model 6). For example, elevated intakes 221 of trans-fatty acids are predicted to be detrimental for liver/kidney dysregulation, but have a 222 minor beneficial effect on leukopoiesis dysregulation. Nevertheless, consuming around 2SD 223 of a-tocopherol above the population mean, while consuming vitamin C and trans-fatty acids 224 at the population mean results in low dysregulation across the four scores, suggesting a 225 systemic benefit of high, but not excessive, a-tocopherol intake (Figure 5 for those that included a-tocopherol intake; 153 combinations were identified for at least 229 one outcome score. Figure 4I shows the distribution of dysregulation scores (all traits) 230 predicted for a 2SD increase in a-tocopherol for any identified effects of micronutrients (after 231 inclusion for potential confounders). For all scores, intakes of a-tocopherol 2SD above our 232 sample mean (mean ± SD = 4.75 ± 2.73) is predicted to lead to reductions in dysregulation or 233 aging Z-scores (below the population average) or changes close to 0 (the population average). 234 235

236
We found statistical support for effects of both macro and micronutrients on liver/kidney 237 function and micronutrient dysregulation. There can be strong covariances among macro and 238 micronutrient intake owing to their co-occurrences in foods (see supplementary materials 239 text S5). For example, vitamin C intake is often considered to be a marker of fruit and 240 vegetable consumption (40), and unsurprisingly in this dataset high vitamin C intake coincides 241 with a high carbohydrate diet (text S5). One may thus question which of these two nutritional 242 levels is a better predictor of dysregulation, or if both must be considered. We evaluated this 243 by fitting the two three-dimensional effects for macro and micronutrients simultaneously 244 (models 7 and 8, excluding and including comorbidities respectively) and assessing model fit 245 relative to corresponding previous models using Akaike information criterion (AIC). For 246 liver/kidney function dysregulation a model including macronutrients is favoured by AIC 247 (Table 2). In contrast, for the micronutrient dysregulation score a model including only 248 micronutrients is favoured by AIC (Table 2). We also note here that even the most complex 249 models fitted, which include both macro and micronutrients, as well as other potential 250 predictors of health status, explain less than 5% of the deviance in these dysregulation scores 251 for this population ( and it is certainly worth testing for interactions between these micronutrients in other 283 epidemiological cohorts. 284 285 We found that consuming a-tocopherol at 2SD above the population average is associated 286 with benefit; in the subject population this level corresponds to 10.21mg/day of a-287 tocopherol. The World Health Organisation recommended intake of a-tocopherol for those 288 aged 65+ is 7.5mg/day in females and 10mg/day in males, thus the value we highlight is not 289 beyond current guidelines (46). Highlighting the importance of considering non-linearity, 290 more extreme intake patterns (e.g., >4 SDs above average) were associated with harm ( Fig.  291   5). This finding accords with the results of RCTs suggesting excessive vitamin E 292 supplementation may increase all-cause mortality (47). We also detected a non-linear effect 293 of carbohydrate intake on dysregulation, which suggests individuals at the upper/lower 294 extremes of the observed carbohydrate intakes suffer poor health. Epidemiological and meta-295 analytic study of the effects of carbohydrates on all-cause mortality in humans has found 296 identical patterns (48). Generally speaking, our study therefore provides further support to 297 the importance of looking beyond 'single nutrient at a time' and monotonic "more is better" 298 analyses (1, 49), to detect interactive effects of nutrients. For space reasons we have not 299 presented and discussed the full range of results for all micronutrient intake interactions 300 detected here. However, we have illustrated an approach that can be used to study the 301 effects of multiple micronutrient intakes on health. The GAM-based approach can either be 302 used for discovery (i.e., as means to detect micronutrient interactions), or as a targeted 303 mechanism-driven paradigm to test specific hypotheses about micronutrient interactions 304 derived from experimental biology or nutrition science (42, 50). Additionally, we encourage 305 readers to explore http://nuage.cohenlab.ca where the effects of micronutrient intakes on 306 dysregulation and aging scores in this population can be visualized interactively, and a priori 307 hypotheses explored. 308 309 Even in our most complex models, the deviance explained remains at around 5%. This is 310 within the bounds of what analyses of other outcomes in this cohort have found (e.g. (51)). 311 Nevertheless, the surfaces we present here show relatively large effects of nutrition on 312 dysregulation levels, which is potentially important as nutrition is readily modifiable. The low 313 deviance explained may be due to a number of factors including: the reporting errors, and 314 systematic reporting biases inherent in epidemiological studies of nutrition (52-54); the fact 315 we consider additive effects of macro and micronutrients rather than interactions (more 316 complex models are theoretically possible but are limited by sample size and visualisation 317 beyond three-dimensions); and the rarity of the extreme dietary profiles at the edges of the 318 surfaces, where the strongest effects are found (the standard error of the surface is a proxy 319 for sample density; e.g. supplementary materials figure S6). An implication of this latter point 320 is that some of the largest effects of nutrient intake we report will only apply to a small 321 proportion of the population, and that our physiology is often robust enough to tolerate 322 relatively wide variation without much consequence. Similar patterns are observed when 323 using the GFN to map evolutionary fitness to organisms that ecologists pre-define as dietary 324 'generalists' (16,(55)(56)(57). This is consistent with an understanding of nutrition in which our 325 ancestors evolved to tolerate an array of dietary patterns (58). Accordingly, homeostasis can 326 be maintained across a wide array of nutritional states, with the caveat that when diet 327 becomes too extreme dysregulation can increase very rapidly ("falling off a dietary cliff"). The 328 tolerance for different diets -the size of the plateau in this analogy -could of course vary as 329 a function of genetic or environmental factors that predispose us to greater risk (59 on an individual's biological profile. Of course, replication in other cohorts is needed to 341 confirm the precise patterns we report, and even then, context-specific clinical 342 recommendations will be essential. geroscience map on to humans living in the community remains unclear. In part, our 361 understanding has been hampered by a lack of techniques that cut across the complexity 362 inherent to nutrition and biological aging in real-world contexts; the approach we present 363 here is a promising start. Future applications could include personalized approaches to aid in 364 healthy aging, and screening of at-risk older adults to ensure they do not fall off the "dietary 365 cliff." Finally, our results advocate against the popular practice of eating to maximize or 366 minimize certain nutrients. The dose-response relationship is often U-shaped, and highly 367 dependent on context (e.g., age, other aspects of diet); targeting in the absence of clear 368 evidence is likely to do more harm than good. were community-dwelling men and women, aged 67-84 years, able to speak English or 377 French, and in good general health at recruitment. Notably, they had to be free of disabilities 378 in activities of daily living, not cognitively impaired and able to walk 300 metres or to climb 379 10 stairs without rest. A structured interview was conducted annually in the NuAge Study at 380 baseline (T1) and for the next 3 years (T2, T3 and T4) to gather the following data. Dietary intake data were collected annually (T1 to T4) using 3 non-consecutive 24-hour diet 400 recalls (71-73). Each set included 2 weekdays and 1 weekend day, with the first administered 401 during the annual face-to-face interview and the others by telephone interviews without prior 402 notice. Based on the USDA 5-step multiple-pass method (74), interviewers recorded a 403 detailed description and portion sizes of all foods and beverages consumed by each 404 participant the day before the interview. Only energy and nutrients coming from foods were 405 considered (i.e., excluding supplements). All interviewers were trained registered dietitians. 406 Energy and nutrient intake were computed from the 24-hour diet recalls using the CANDAT- to their availability in NuAge. In total 30 biomarkers were used to calculate dysregulation 414 globally and for five systems that a previous study validated as being largely independent (10): 415 1) Oxygen transport; 2) Liver/kidney function; 3) Leukopoiesis; 4) Micronutrients; and 5) 416 Lipids (Table 1) Exploratory analysis of other datasets in our possession indicated the discrepancy was likely 450 due to the limited age range rather than the absence of CRP (data not shown). Calculation of 451 biological age was based on previous work by Levine (35). We first searched for biomarkers 452 that correlated with chronological age and found 13 with r > 0.1 (p £ 0.05). After removing 453 three biomarkers with numerous missing values (non-HDL cholesterol, LDL, and estimated 454 glomerular filtration rate), we were left with the following list: hemoglobin, hematocrit, red 455 blood cell count, red cell distribution width, monocyte count, albumin, folate, creatinine, 456 blood urea nitrogen, and lymphocyte percentage. Using these biomarkers, we calculated 457 biological age as previously described (35,78). We first tested the hypothesis that increased protein is associated with improved health in 474 old age. We estimated the effects of daily intake of macronutrients (protein, carbohydrate 475 and fat, in kJ) on each dysregulation/aging score using GAMs. In all models, 476 dysregulation/aging score at an observation was the outcome. In model 1, macronutrient 477 intakes at those observations were fitted as a three-dimensional smooth term (thin-plate 478 spline), and individual subject ID was fitted as a random effect (there are multiple 479 observations per individual). However, the dataset is made up of individuals whose energy 480 requirements likely vary due to height, weight, age, sex and physical activity level. Thus, we 481 also estimated macronutrient intakes relative to what is typical for an individual in this 482 population given the aforementioned variables (estimated from the residuals of a model of 483 intake; see supplementary materials text S4). In model 2, we modelled dysregulation scores 484 as a function of relative macronutrient intake using a GAM as above. 485

486
In model 3, we tested for the effects of confounders by refitting model 2 to include potential 487 confounders as additive effects; sex (men / women), smoking status (current / not current), 488 current income, number of years of education, alcohol intake (g/d), physical activity level 489 (PASE) and age were included. We ran a separate model with comorbidities as an additional 490 potential confounder (model 4). Numeric predictors were Z-transformed prior to fitting and 491 were included in the model as smooth terms, and categorical predictors as parametric terms. 492

493
In the second part of our analyses, we considered the effects of those micronutrients for 494 which intake data were available. Micronutrient intakes are likely to be highly correlated. 495 Therefore, we performed hierarchical clustering on correlations of micronutrient intakes 496 based on correlation distance using the 'hclust' function in base R. For any clusters of highly 497 correlated micronutrients we performed a principal component analysis (PCA; 'prcomp' 498 function in base R) and used the first principal component (PC1) as our measure of intake for 499 the nutrients within that cluster. The number of micronutrients adds complexity to 500 understating their interactive effects because: 1) the estimation of multidimensional smooth 501 terms in GAMs becomes challenging as the number of dimensions grows, and 2) the smooth 502 terms in GAMs must be interpreted visually. For these reasons we restricted ourselves to 503 considering 3 micronutrient dimensions at a time. For each three-way combination of 504 micronutrient intake (PC1 for clusters) we fitted a GAM with a three-dimensional smooth 505 term for the three micronutrients and a random effect for individual. We then identified those 506 models with significant effects after correction for the false discovery rate (FDR, q < 0.05; 507 (81)). We note that correction by FDR assumes that p-values are independent, which is not 508 the case here. However, overlooking this non-independence is conservative in that it will 509 result in fewer significant effects rather than more, and thus we proceeded without 510 correction. In the main text we interpret specific cases of interest, however, all results can be 511 accessed at http://nuage.cohenlab.ca. We also tested for effects of micronutrients alongside 512 correction for the potential confounders discussed above (models 5 and 6), and by modelling 513 micro and macronutrient intakes simultaneously (models 7 and 8). Note that including two 514 three-dimensional smooth terms (one for micronutrients and one for macronutrients) is not 515 the same as a six-dimensional smooth term; this approach is substantially less power-hungry 516 but still allows us to adjust micronutrient models for macronutrient intake, and vice-versa. 517 See supplementary materials (text S1) for a concise list of all models implemented. The main text contains a complete cases analysis where we report results from analyses on 529 all observations for which relevant predictor and dysregulation score data were available (i.e. 530 inclusive dataset with missing income data excluded). Supplementary texts S2 and S3 contain 531 the results of two sets of sensitivity analyses. In the first we have imputed missing income 532 data (31%) as the participant-specific mean value to increase our sample size. In the second 533 we have analysed a subset of the data (<50% of the total data) where we have excluded any 534 observation at which the subject was recorded as either; diabetic (type-1 or 2), reported as 535 being on a medically prescribed diet, having a BMI outside the range of 22 to 29.9, or that 536 came from a subject with a coefficient of variation of weight > 0.04 over the course of all the 537 observations (i.e. for whom weight fluctuated substantially).             which is as in model 1, but the three-dimensional macronutrient smooth term is now 29 relative intake of each macronutrient at the ith observation (estimated as described in text 30 S4 'Estimation of Relative Intake'). 31 32 Model 3. 33 which is as in model 3 but includes a further smooth term, x7, which is number of 44 comorbidities of the individual at observation i.
which is as in model 3 but considers a combination of three micronutrients (or PC1 of 55 clusters thereof), f(k1i, k2i, k3i), determined to be of interest following the fitting of 56 micronutrient-specific models. 57 58 Model 6. 59 which is as in model 6 but includes the seventh smoothed term for number of comorbidities 61 (x7). 62 63 Model 7. 64 which is as in model 3 but with f7(k1i, k2i, k3i), which is an additional three-dimensional 67 smooth term for the intake of a combination of three micronutrients (or PC1 of clusters 68 thereof) determined to be of interest following the fitting of micronutrient-specific models. We re-ran our analyses on a subset of the data in which we had excluded any observations 103 where the individual was recorded as any of the following: diabetic (type-1 or 2), reported 104 as being on a medically prescribed diet, having a BMI outside the range of 22 to 29.9, or 105 came from a subject that had a coefficient of variation of weight > 0.04 over the course of 106 all the observations within the dataset. The logic being that, in doing so we would be 107 ensuring that our results were not driven by individuals who had a dietary associated 108 chronic disease and extreme dietary profile that was disproportionately affecting our 109 models. Table S1 shows the profiles of the populations captured within the different 110 datasets. These restrictions excluded over the half of the data, which substantially reduced 111 statistical power. With this reduced dataset our analyses estimated effects with the same 112 sign as those in the main text, but for macronutrients (and liver/kidney function 113 dysregulation with micronutrients) failed to reach the level of statistical significance. 114 115 Figure S3 displays the same results as figure 2 in the main text. For this restricted dataset 116 the models estimated the same effects as those for the inclusive dataset, although the 117 effects for liver/kidney function dysregulation, micronutrient dysregulation and biological 118 age score were non-significant. Figure S4 shows the effects of a-tocopherol, vitamin C and 119 trans-fatty acid intake on leukopoiesis, liver/kidney function, micronutrient and global 120 dysregulation score. Here all potential confounding variables have been taken into account 121 (i.e. model 6). We see the qualitatively similar effects to those presented in the main text, 122 although the effects are non-significant for all cases but global dysregulation score. individuals with a relative intake value of 100, eat 100% more of that macronutrient per day 137 (in kJ) than is predicted to be typical for this population given their age, sex, weight, height 138 and level of physical activity. Conversely, individuals with a relative intake value of 0 eat 139 what is the expected given these factors weight, height age, sex and PASE score. 140

141
We tested for associations between the composition of the diet in terms of percentage 142 energy from the three macronutrients, protein, carbohydrates and lipids and intake of the 143 micronutrients (or PC1 for clusters) considered. We used mixture models (MMs), also 144 known as Scheffe's polynomials (1), where micronutrient intake (z-transformed to one SD) 145 was the outcome and percentage energy from each macronutrient were the predictors. For 146 each micronutrient we fitted 5 MMs; MM 1 was a null model and MMs 2 through 5 tests for 147 increasingly complex linear through non-linear effects of diet composition on micronutrient 148 intake (see equations 1 through 4 in (1)). It is notable that MM 2 is identical to the 149 substitution models commonly used in nutritional epidemiology (2, 3). We compared among 150 MMs using Akaike Information Criterion (AIC; (4)), assuming that the simplest model within 151 2 AIC points of the lowest AIC observed was best supported. MMs were implemented using 152 the mixexp package in R (1). For micronutrients of interest, we visualised the predictions for 153 the AIC-supported MM using right-angle mixture triangles (RMTs; (5)). For all micronutrients 154 AIC favoured models other than the null model (table S11), suggesting that dietary 155 macronutrient composition is associated with intake of micronutrients. Figure S5 Table S1:       Table S7: Numbers of models with significant 3-way smooth terms for effects of 197 micronutrients on different combinations of outcomes (micronutrient-specific models); 1 198 indicates the outcome score is in the group (i.e. 1, 0, 0, 0, 0, 0 indicates leukopoiesis 199 dysregulation alone, where 1, 1, 1, 1, 1, 1 indicates all scores).

211
Table S10: Akaike information criterion (AIC) and deviance explained (%) by models 3, 5, and 7, applied to the inclusive data set with missing 212 income data excluded, for liver/kidney and micronutrients dysregulation. These models contained, alcohol intake, sex, smoking status, income, 213 education level, age, and physical activity (PASE) alongside the nutritional predictors stated. 214 215 and 0.8 (conventionally considered an effect of large biological magnitude (6)). Individuals 235 with a relative intake value of 100, eat 100% more of that macronutrient per day (in kJ) than 236 is predicted to be typical for the population given their age, sex, weight, height and level of 237 physical activity level. Conversely, individuals with a relative intake value of 0 would eat the 238 required amount of that macronutrient per day.  Figure S6. Effects with corresponding standard errors (se) of relative macronutrient intake 294 (relative to estimated requirements; see text S1) on liver/kidney function dysregulation 295 score as predicted by model 4 applied to the inclusive dataset with imputation of missing 296 income data. Left hand surfaces show effects of protein (x-axis), and carbohydrate (y-axis) 297 intake, middle show protein and lipid, and the right is carbohydrate and lipid. The third 298 macronutrient is held at the values given on all panels. Warm colours indicate high 299 dysregulation, and cool colours low dysregulation. All scores were Z-transformed to one SD, 300 and surfaces colours are scaled such that deep blue and red represent effects of at least -0.8 301 and 0.8 (conventionally considered an effect of large biological magnitude (6)). In all cases 302 predictions assume the macronutrient not displayed on either the x-or y-axis is held at the 303 population median. Numeric confounding variables included in model 4 were alcohol intake, 304 income, education level, age, physical activity levels (PASE) and number of comorbidities, 305 and predictions assume population mean values. Predictions are for men and assume a non-306 smoker. 307