Towards causal modeling of nutritional outcomes

Ksenia Gasnikova, Olivier Allais, Michèle Sebag
Proceedings of The 2021 Causal Analysis Workshop Series, PMLR 160:5-19, 2021.

Abstract

This paper aims at observational causal modelling, investigating the causal relationships between food consumption and health status, exploiting the proprietary Kantar database. This database describes the socioeconomic characteristics and consumption habits of a few dozen thousands households; in particular, the consumed food items are documented almost at the level of precision of barcodes. A first challenge for this observational causal study lies in the number of hidden confounders, ranging from genetic factors to life styles (i.e. smoking and sport habits), not documented in the data. Taking inspiration from the Deconfounder approach (Wang and Blei, 2019b), substitute hidden confounders based on dietary patterns − viewed as characteristics of the alimentary lifestyle − are extracted from the database and exploited to block the biases due to hidden confounders. A second challenge lies in the fact that the data size hardly allows for investigating a number of fine-grained interventions. We thus define a new type of intervention, enabled by the data structure and referred to as macro-intervention, acting on the full basket of food items; an example of such macro-intervention is to replace every non-organic product in a household basket with its organic counterpart. The average treatment effect of this macro-intervention is assessed in the context of the substitute hidden confounders, using inverse propensity weighted estimates to control for covariates such as wealth or education.

Cite this Paper

BibTeX
@InProceedings{pmlr-v160-gasnikova21a, title = {Towards causal modeling of nutritional outcomes}, author = {Gasnikova, Ksenia and Allais, Olivier and Sebag, Mich\`ele}, booktitle = {Proceedings of The 2021 Causal Analysis Workshop Series}, pages = {5--19}, year = {2021}, editor = {Ma, Sisi and Kummerfeld, Erich}, volume = {160}, series = {Proceedings of Machine Learning Research}, month = {16 Jul}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v160/gasnikova21a/gasnikova21a.pdf}, url = {https://proceedings.mlr.press/v160/gasnikova21a.html}, abstract = {This paper aims at observational causal modelling, investigating the causal relationships between food consumption and health status, exploiting the proprietary Kantar database. This database describes the socioeconomic characteristics and consumption habits of a few dozen thousands households; in particular, the consumed food items are documented almost at the level of precision of barcodes. A first challenge for this observational causal study lies in the number of hidden confounders, ranging from genetic factors to life styles (i.e. smoking and sport habits), not documented in the data. Taking inspiration from the Deconfounder approach (Wang and Blei, 2019b), substitute hidden confounders based on dietary patterns − viewed as characteristics of the alimentary lifestyle − are extracted from the database and exploited to block the biases due to hidden confounders. A second challenge lies in the fact that the data size hardly allows for investigating a number of fine-grained interventions. We thus define a new type of intervention, enabled by the data structure and referred to as macro-intervention, acting on the full basket of food items; an example of such macro-intervention is to replace every non-organic product in a household basket with its organic counterpart. The average treatment effect of this macro-intervention is assessed in the context of the substitute hidden confounders, using inverse propensity weighted estimates to control for covariates such as wealth or education.} }
Endnote
%0 Conference Paper %T Towards causal modeling of nutritional outcomes %A Ksenia Gasnikova %A Olivier Allais %A Michèle Sebag %B Proceedings of The 2021 Causal Analysis Workshop Series %C Proceedings of Machine Learning Research %D 2021 %E Sisi Ma %E Erich Kummerfeld %F pmlr-v160-gasnikova21a %I PMLR %P 5--19 %U https://proceedings.mlr.press/v160/gasnikova21a.html %V 160 %X This paper aims at observational causal modelling, investigating the causal relationships between food consumption and health status, exploiting the proprietary Kantar database. This database describes the socioeconomic characteristics and consumption habits of a few dozen thousands households; in particular, the consumed food items are documented almost at the level of precision of barcodes. A first challenge for this observational causal study lies in the number of hidden confounders, ranging from genetic factors to life styles (i.e. smoking and sport habits), not documented in the data. Taking inspiration from the Deconfounder approach (Wang and Blei, 2019b), substitute hidden confounders based on dietary patterns − viewed as characteristics of the alimentary lifestyle − are extracted from the database and exploited to block the biases due to hidden confounders. A second challenge lies in the fact that the data size hardly allows for investigating a number of fine-grained interventions. We thus define a new type of intervention, enabled by the data structure and referred to as macro-intervention, acting on the full basket of food items; an example of such macro-intervention is to replace every non-organic product in a household basket with its organic counterpart. The average treatment effect of this macro-intervention is assessed in the context of the substitute hidden confounders, using inverse propensity weighted estimates to control for covariates such as wealth or education.
APA
Gasnikova, K., Allais, O. & Sebag, M.. (2021). Towards causal modeling of nutritional outcomes. Proceedings of The 2021 Causal Analysis Workshop Series, in Proceedings of Machine Learning Research 160:5-19 Available from https://proceedings.mlr.press/v160/gasnikova21a.html.

Related Material