Task 2.1. Innovative characterization of the diet and type 2 diabetes risk and interactions with the socioeconomic environment (PI: G. Fagherazzi, CESP)

Aims. First, we will evaluate associations between diets, characterized in an innovative way (energy and nutrient intakes distribution over the day, snacking frequency, diet quality index, diet diversity index, total dietary antioxidant capacity, dietary acid load, life-course alcohol intake, cumulative consumption of artificial sweeteners, dietary patterns derived from latent class analysis…), and T2D risk. Second, we will conduct complex analyses of the associations between socioeconomic factors (age, salary, level of education, number of children, age at first delivery, deprivation index FDep (Rey et al., 2009), size of the living and working area, retirement), diets (focusing on dietary patterns ans diet quality indexes), and T2D risk. At month 12, we will be able to provide the first set of results on the association between diet, socioeconomic factors and T2D risk. We will focus on dietary patterns and diet quality indexes. At month 36, we will have a complete overview of the relationships between diet and diabetes risk, stratified by various socioeconomic factors.
Methods and data. We will use data from the large prospective E3N cohort study (www.e3n.fr) of 98 995 female teachers, initiated in 1990. Participants returned mailed questionnaires to update information on lifestyle factors and newly diagnosed diseases every 2 to 3 years. Dietary data were collected in 1993 and 2005 using a validated diet history questionnaire. The consumed frequencies and quantities consumed of 208 food items, over the past year, were reported for eight meals during the day, from breakfast to after-dinner snacks (including aperitifs, before lunch and dinner).
Expected results. We will have a deeper understanding of the interactions between various socioeconomic factors and dietary factors/patterns. We will be able to provide very specific identity cards of high-risk individuals (for example “women with high intakes of fat and artificial sweeteners, living in a suburban area, and with a high deprivation index and a low level of education are those with the highest risk of developing type 2 diabetes”) and quantify the risks of T2D associated with these profiles. The results of the task will directly help to target high risk individuals for an intervention and for tailored nutritional recommendations.
Risks. Data are already available and ready for analyses and we benefit from a long follow-up (26 years) and large statistical power (more than 5 000 T2D cases).
Collaborations. The characterization of relevant socioeconomic factors will be conducted in collaboration with the other tasks of this WP. In addition, the at-risk-dietary patterns identified in the E3N cohort (women with a limited range of social categories) will be used to identify at-risk individuals in the population-wide Kantar dataset in Tasks 2.2 and 2.3 and results will be pooled with those of 2.4.

Task 2.2. Substituability of (deleterious) dietary patterns (PI: O. Allais (INRA-Aliss), M. Sebag (INRIA-CNRS)

Aims. An economic model has been previously designed to estimate the loss of (short-term) welfare induced by the adoption of nutritional recommendations by consumers, depending on consumers characteristics (Irz et al., 2015). The general principle is to estimate the compensating value (the additional budget) that would be necessary to give a consumer to comply with one recommendation (e.g. increase the F&V consumption) without any loss of utility. In this Task, the novelty is to consider this cost estimation at the integrated level of dietary patterns, in order to support more acceptable and socio-economically-aware recommendations. The goal is twofold: identifying dietary patterns in relation with T2D risk; identifying whether and how a potentially deleterious dietary pattern can be substituted with another (less deleterious) one in a socio-economically acceptable way for consumers.
Methods and data. The study will be conducted based on the Kantar Worldpanel dataset, describing the food purchases of a consumer panel including 27,291 representative French households together with a socioeconomic description of the household (e.g., age, gender, BMI of each household member) over 10 years. A first milestone is to infer the dietary patterns at the individual level by disaggregating the household patterns. Existing disaggregation methods (Chescher, 1998; Allais and Tressou, 2009) will be extended using Machine Learning approaches, specifically the Multiple Instance setting Almost (Patrini et al., 2015). The second milestone is to determine comprehensive individual clusters, providing an integrated perspective on socio-economic and dietary patterns, and their relationships; these clusters will be based on a compressed representation to account for the diversity of the purchase description, using linear or non-linear dimensionality reduction techniques (Bengio et al., 2013). The third milestone is to assess the T2D risk value of each dietary pattern, based on the relative risks associated to each food group. The fourth milestone is concerned with defining a "metric" in the dietary pattern landscape. Basically, our goal is to be able to emit recommendations for a targeted consumer group, such that the recommended dietary pattern has lesser T2D risk while minimizing the loss of utility. In other words, the recommendation task defines a multi-objective optimization problem, maximizing both the chance for the recommendation to be accepted, and the positive impact on the consumer health or well-being. The proposed method, learning a metric on the dietary pattern landscape, will leverage the recommendation approaches used in e-commerce and the collaborative filtering techniques (two products are "close" if they are consumed by similar people; the consumption metric will be refined using the prior knowledge related to the food groups), and the economic approach of consumer utility.
Risk. The multi-objective strategy opens a range of possibilities: from recommending a dietary pattern very similar to an actual one and only marginally better T2D-wise, to recommending a dietary pattern rather different and much healthier. The optimal trade-off needs be determined in an empirical way.
Collaboration. Task T2.2 is a multi-disciplinary task, requiring the tight collaboration of economics, and machine learners. As suggested by Bajari et al. (2015), the connexion between econometrics and machine learning methods can open new avenues in demand estimation.

Task 2.3. Lifecycle events and variation of T2D risk: a causal inference analysis (PIs; O. Allais (INRA Aliss), G. Fagherazzi (INSERM CESP))

Aims. The retirement effect on consumption expenditure has been already largely studies (Aguiar and Hurst, 2008; Aguila and al. 2011; Li and al. 2015). However, these studies focused their analysis on aggregate non-durable consumptions. The quality and variety dimensions of dietary patterns were absent from their analysis. It is also the case for studies on the impact of divorce, unemployment, household move, and childbirth events. Yet, these dimensions are crucial to understand the effects of lifecycle events on social inequalities and individual health status evolution over life course. The main objective of this task is to characterize individual groups with the highest T2D risk over lifecycle using their dietary patterns. We will achieve our goal by measuring how much lifecycle events such as retirement, divorce, unemployment, household move, and childbirth events alter (strengthen or reduce) individual’s consumptions of food groups affecting T2D risk.
Method and data. Our objective will be achieved by using regression discontinuity approach to estimate the causal impact of each life cycle event on consumptions of deleterious and healthy food product. Then, we will develop an epidemiological model to measure the impact of food consumption changes caused by each life cycle events on T2D incidence and health status evolution. A panel of almost 4,000 households over 8 years from Kantar scanner data will be employed. Expected results. We expect that life cycle event involving a drop in income would decrease diet quality.
Risks. One issue of this analysis will be to disentangle the income variation effect from the effect of variation in available time. The detailed data on shopping trips will be used to overcome this difficulty. Another issue will be to estimate the “true” causal impact of each event. Obtaining it supposes the exogeneity of life cycle events on food consumption decisions. This assumption can strongly be questioned for each event. Instrumental variable approaches will be used to overcome the risk.
Collaborations. It is a joint work between Aliss and CESP. Aliss will conduct all causal estimations and CESP will help building the epidemiological model to assess the health impact of each life cycle event.

Task 2.4. Social determinants of dietary patterns and health-related behavior change conditions (PIs: F. Régnier (INRA Aliss), C. Licoppe (TelecomParisTech))

Aims. Therapeutic compliance as well as dietary adherence involves socio-demographic factors, the nature of the medication / dietary prescriptions, the doctor-patient relationship (Fainzang, 2001, 2006), dietary models and norms (Fournier, 2011). In the field of diabetes, studies most often examine compliance factors among patients with diabetes (Krass et al., 2015; Sapkota et al., 2015) but very few of them deal with a prevention prospect, and the precise knowledge of the conditions under which health behaviors can be modified.
We will investigate the social determinants of current dietary patterns and determine the conditions for health-related behavior modification for at-risk people. For this, we will assess the joint effects of: socio-demographic profile (social status, integration and socialization) and the related constraints (budget, time, family obligations, etc.), with a dynamic dimension (place in the life cycle, social trajectory); lifestyle and health variables; food preferences and health perceptions (health concerns vs. pleasure, short term vs. long term); individual motivation factors (achievement orientation or taste for competition vs. desire for social integration, concern for one’s appearance or health vs. priority given to others, e.g. children).
Methods and data. First (M1 to M18), in order to assess how social factors determine resistance to change, we will conduct a field survey with 50 interviews of a comparative sample: with or without specific medical follow-up care; in a socially diverse sample, to evaluate the role of social belonging in diabetes risk management and compliance with recommendations. The body of interviews will undergo classic content analysis and lexical analysis. Second (M18 to M24), in order to characterize which dietary recommendations may be adopted at lower cost, we will organize six to eight focus groups of approximately 10 at-risk people and association members, of diverse social profiles, discussing three or four of the major results of the field survey. These semi-structured group interviews, whose methodological rules ensure scientific validity, are especially interesting to grasp reactions to and opinions on a few key points, and gather participants’ opinions and expectations of an intervention.
Expected Results. Identification of economic, social and cultural conditions for, of levers for and hindrances to the adoption of preventive behaviors; assessment of perceived benefits and costs by at-risk people to modify health behaviors; identification of tailored recommendations suited to the profile of at-risk people depending on their social characteristics.
Risks. The methodology has been validated in previous research. A few difficulties may be pointed out: possible sampling bias, overcome by the complementarity of recruitment sites (medical milieu/associative milieu), and by controlling for individual social diversity.
Collaboration. This task will be conducted in close collaboration with Tasks 2.1-3 in order to recruit people from to the same at-risk groups in our study population.