1. Reduced Rank Regression – a powerful statistical method for identifying empirical dietary patterns
Gina Ambrosini PhD
Senior Research Scientist
MRC Human Nutrition Research, Cambridge
EUCCONET International Workshop, Bristol October 2011

2. Why dietary patterns ?
The human diet is complex – we do not eat nutrients or foods in isolation
Single food/nutrient studies are frequently null e.g. fat intake and obesity; these do not consider total dietary intake
Strong co-linearity between dietary variables; ; difficult to separate effects, may be too small to detect
Numerous dietary variables (foods & nutrients) lead to too many statistical tests
Studies of dietary patterns i.e. combinations of total food intake can overcome many of these problems
Studies of single foods or nutrients ignore the complexity of the human diet; synergism

3. What nutrition epidemiologists want to know …
Reduced Rank Regression ? ?
Disease or
Health Outcome
Dietary
Pattern
PCA or Factor Analysis
Empirical versus other patterns; PCA and cluster methods are exploratory, depend on study population covariance matrix, therefore reproducbility in different populations unknown; not hypothesis based or disease specfici.
PCA, factor and cluster are exploratory methods; depend on study population and covariance matrix
Study specific: reproducibility unknown in different populations
Explain variation in food intakes but not necessarily nutrients – the end product of diet
Clusters are based on membership; non-quantitative
Not disease-specific or hypothesis-based
Cluster Analysis
Dietary Indices
Eg. Healthy Eating Index

4. Empirical Dietary Patterns
E.g. Principal Components Analysis (PCA), Factor Analysis and Cluster Analysis
Data reduction techniques; identify latent constructs in data = patterns
Take advantage of co-linearity
Consider total diet; ‘real-life’ consumption and synergism
Produce uncorrelated dietary patterns (or clusters) suitable for multivariate models
Exploratory, data-driven, study specific: reproducibility unknown in different populations
Explain variation in food intakes but not necessarily nutrients – the end product of diet
Not disease-specific or hypothesis-based
Food Intakes
These methods have been most frequently used in the published literature
Dietary Patterns

5. Reduced Rank Regression – a novel empirical approach
In 2004 … researchers in Germany … presented

6. Reduced Rank Regression (RRR)
A hypothesis-based empirical method for identifying dietary patterns
Similar to PCA and factor analysis but requires a 2nd set of data = response variables
Response variables should be on the pathway between food intake and outcome of interest
RRR dietary patterns are linear combinations of food intake
that explain the maximum variation in a set of response variables
Dietary Pattern
So, let me walk you through this …
As nutrition scientists, we expect that the food we eat contains particular nutrients and non-nutrient active ingredients that have health effects in the body. We ingest food, from which nutrients are absorbed and have different effects in the body to influence disease risk.
We know that for example, saturated fats sourced from animal-based foods eg. butter, milk, meats, contribute to the development of calcified plaques in blood vessels and ultimately arteriosclerosis…
Now imagine that we can identify dietary patterns or combinations of food that are specifically linked to selected nutrient intakes or even biomarkers that are susceptible to dietary intake e.g blood cholesterol
Response variables should be intermediates on the biological pathway between predictors and health outcome of interest
Disease or
Outcome
of Interest
Food
Intake
Nutrients Or
Biomarkers
Predictors
Responses

7. Example - ALSPAC
Measured dietary intake using a 3d food diary at 7, 10 and 13 years of age
We hypothesised that: a dietary pattern that could explain the variation in dietary energy density, % energy from fat, and fibre at 7, 10 and 13 y would be prospectively assoc with body fatness measured at 9, 11, 13, 15 y

8. Each dietary pattern is a linear combination of weighted food intakes
Example RRR - ALSPAC
1st Dietary Pattern:
Energy-dense,
high in fat,
low in fibre
Predictors
Food Group Intakes
Responses
Nutrient Intakes
Dietary
Pattern 1
Fat
Fruit
Veg
Dietary
Pattern 2
3-day food
diary F3 F4
OBESITY(fat mass)
Fibre F5 F6
Energy
Density
Dietary
Pattern 3 F7
F8…
Explain what dietary pattern z-score means … level of adherence to dietary pattern
Each dietary pattern is a linear combination of weighted food intakes
that explains the max variation in ALL response variables -1st pattern often explains the most
Such that for each dietary pattern a z-score is calculated as
= W1(Food1 Intake) + W2(Food2 Intake) + W3(Food3 Intake) + …

9. ALSPAC energy-dense, high fat, low fibre dietary pattern
This pattern was virtually identical when we conducted RRR at 7 years of age …

10. Girls
 Age
9 y
11 y
13 y
15 y
 Dietary Pattern
n=2868
n=2274
n=2007
n=1556
7 y
0.08
0.07
(95% CI)
( )
( )
( )
( )
p-value
<.0001
<0.001
10 y
0.05
0.04
( )
( )
0.01
0.02
13Y
-0.01
( )
0.68
Boys
n=2854
n=2118
n=1863
n=1345
0.09
0.06
( )
0.012
0.006
( )
( )
0.65
0.64
( )
0.45
ALSPAC – change in Fat Mass Index (z-score) with a SD increase in energy-dense, high fat, low fibre dietary pattern z-score
These results are based on linear regression using completers only, i.e. only subjects who provided data at all relevant time points – longitudinal analyses utilising more subjects are currently underway and are showing similar relationships
Adjusted for age at fat mass assessment, dietary misreporting, physical activity (cpm)

11. Cross-cohort comparisons: ALSPAC v Raine Study
PhD project – Geeta Appannah University of Cambridge and MRC Human Nutrition Research:
An almost identical energy-dense, high fat, low fibre dietary pattern seen at 14 and 17 y in The Western Australian Pregnancy Cohort (Raine) Study, a contemporaneous birth cohort.
Similar factor loadings for an energy-dense, high fat, low fibre dietary pattern in a FFQ and a food diary at 14 y of age in the Raine Study
Geeta Appannah, MRC Human Nutrition Research

12. Comparisons of RRR and PCA patterns
Study
RRR response variables
Outcome
Multi-Ethnic Study of Atherosclerosis (US)
CRP, IL-6, Fibrinogen, Homocysteine
Sub-clinical atherosclerosis
EPIC Potsdam (Germany)
Fibre, Magnesium, alcohol
Type 2 Diabetes
% Energy from saturated fat, PUFA, MUFA, protein and carbohydrate
All cause mortality
SFA, MUFA, n-3 PUFA, n-6 PUFA
Breast cancer incidence
Tehran Lipids and Glucose Study
Total fat, PUFA/sat fat, cholesterol, fibre, calcium
Obesity
Although the PCA and RRR patterns in these studies had similar nutrient profiles; these studies reported stronger associations between RRR-based dietary patterns and outcomes
RRR patterns explain more variation in the response variables
Gina Ambrosini

13. Caution - using biomarkers as response variables
Biomarkers as response variables should be chosen carefully:
So they are true intermediates and not a proxy for the outcome of interest
Should be on pathway;
Therefore must be susceptible to dietary intake – relevant to more novel biomarkers
Dietary Pattern
Diabetes
Food
Intake
Blood Glucose
Insulin Resist.
Predictors
Responses
Gina Ambrosini

14. Generalisability of RRR patterns
Imamura et al (2010) applied RRR dietary patterns that were associated with type 2 diabetes in three different cohorts to the Framingham Offspring Study
All patterns were characterised by high intakes of meat products, refined grains and soft drinks
Dietary Pattern
RRR response variables
Risk of T2D in Framingham Offspring Study
EPIC Potsdam (Germany)
Fibre, Magnesium, alcohol
1.14 (0.99 – 1.32)
Nurses Health Study (US)
Inflammatory markers
1.44 (1.25 – 1.66)
Whitehall II (UK)
Insulin resistance *
1.16 (1.00 – 1.35)
Imamura F et al. Generalizability of dietary patterns associated with type 2 diabetes mellitus. AJCN 2010; 90(4):
Gina Ambrosini

15. Limitations RRR appears to be a robust and powerful method, however:
Reproducibility, generalisability of patterns – only 1 published study
RRR depends on existing knowledge in order to choose response variables
Response variables must be chosen very carefully to avoid circular analysis
Biomarkers as response variables: must be an intermediate and not a proxy for the outcome/disease
If response variables are not intermediates, then false relationships may be observed b/w dietary pattern and health outcome, as response variable is a proxy for outcome
Gina Ambrosini

16. Acknowledgements Funding from:
Dr Pauline Emmett, Dr Kate Northstone, & the ALSPAC Study Team
Ms Geeta Appannah, PhD scholar, MRC Human Nutrition Research
Mr David Johns, PhD scholar, MRC Human Nutrition Research
Dr Anna Karin Lindroos, Swedish Food Authority, Uppsala (prev. HNR)
Funding from:

17. MRC Human Nutrition Research
Cambridge, UK

18. Reported Associations with Other RRR Dietary Patterns
Study
RRR response variables
Outcome
Multi-Ethnic Study of Atherosclerosis (US)
CRP, IL-6, Fibrinogen, Homocysteine
Sub-clinical atherosclerosis
Insulin Resistance Atherosclerosis Study (US multi-ethnic cohort)
Plasminogen activator inhibitor 1, Fibrinogen
Carotid artery atherosclerosis (IMT, CAC)
Coronary Risk Factors for Atherosclerosis in Women (CORA) Germany
LDL and HDL cholesterol lipoprotein (a)
CRP, C-peptide (insulin resist)
Coronary artery disease
Nurses Health Study (US)
Inflammatory markers
Type 2 Diabetes
Framingham Offspring Study (US)
BMI, fasting HDL-C, TG, glucose, hypertension (BP residuals)
EPIC Potsdam (Germany)
Fibre, Magnesium, alcohol
% Energy from saturated fat, PUFA, MUFA, protein and carbohydrate
All cause mortality
SFA, MUFA, n-3 PUFA, n-6 PUFA
Breast cancer incidence
Tehran Lipids and Glucose Study
Total fat, PUFA/sat fat, cholesterol, fibre, calcium
Obesity
ALSPAC
Energy density
% energy from fat
Fibre density
Child obesity at 7, 9, 11, 13, 15y
The studies highlighted in red have tested PCA-derived dietary patterns as well and reported stronger associations between RRR-based dietary patterns and their health outcome
Gina Ambrosini