Studying dietary intake in daily life through multilevel two-part modelling: a novel analytical approach and its practical application

Abstract

Background Understanding which factors influence dietary intake, particularly in daily life, is crucial given the impact diet has on physical as well as mental health. However, a factor might influence whether but not how much an individual eats and vice versa or a factor’s importance may differ across these two facets. Distinguishing between these two facets, hence, studying dietary intake as a dual process is conceptually promising and not only allows further insights, but also solves a statistical issue. When assessing the association between a predictor (e.g. momentary affect) and subsequent dietary intake in daily life through ecological momentary assessment (EMA), the outcome variable (e.g. energy intake within a predefined time-interval) is semicontinuous. That is, one part is equal to zero (i.e. no dietary intake occurred) and the other contains right-skewed positive values (i.e. dietary intake occurred, but often only small amounts are consumed). However, linear multilevel modelling which is commonly used for EMA data to account for repeated measures within individuals cannot be applied to semicontinuous outcomes. A highly informative statistical approach for semicontinuous outcomes is multilevel two-part modelling which treats the outcome as generated by a dual process, combining a multilevel logistic/probit regression for zeros and a multilevel (generalized) linear regression for nonzero values. Methods A multilevel two-part model combining a multilevel logistic regression to predict whether an individual eats and a multilevel gamma regression to predict how much is eaten, if an individual eats, is proposed. Its general implementation in R, a widely used and freely available statistical software, using the R-package brms is described. To illustrate its practical application, the analytical approach is applied exemplary to data from the Eat2beNICE-APPetite-study. Results Results highlight that the proposed multilevel two-part model reveals process-specific associations which cannot be detected through traditional multilevel modelling. Conclusions This paper is the first to introduce multilevel two-part modelling as a novel analytical approach to study dietary intake in daily life. Studying dietary intake through multilevel two-part modelling is conceptually as well as methodologically promising. Findings can be translated to tailored nutritional interventions targeting either the occurrence or the amount of dietary intake.

Individual differences in the dietary response to stress in ecological momentary assessment: Does the individual‐difference model need expansion?

“…These models were run using the R-package brms (Bürkner, 2017(Bürkner, , 2018, which supports Bayesian multilevel modelling. Details on this type of analysis (e.g., implementation and interpretation) can be found in Ruf, Neubauer, et al (2021).…”

“…We found fairly strong cross-part correlations in the multilevel two-part models indicating that individuals, who consume on average more energy when they eat, eat less often. Not accounting for this (i.e., running separate models) can cause bias particularly in the continuous part, as higher values of food intake will be underrepresented and smaller values overrepresented (see Ruf, Neubauer, et al, 2021, for a detailed description of this problem). This bias is still present when one is only interested in the continuous part and therefore choses to fit a single model (Su et al, 2009).…”

“…The distinct findings regarding the occurrence and the amount of food intake as well as the strong cross‐part correlations in the present study highlight the importance of multilevel two‐part models when examining the stress and eating relationship in daily life. For this reason, future studies should incorporate multilevel‐two part modelling (practical guidance on this type of analysis can be found in Ruf, Neubauer, et al, 2021).…”

“…Data were collected within the Eat2beNICE‐APPetite‐study (parts of the data have been used to study different research questions, see Ruf, Koch, et al, 2021, and Ruf, Neubauer, et al, 2021). Participants completed two in‐person sessions as well as an EMA period.…”

“…This type of data can be challenging as traditional linear multilevel modelling cannot be applied. A promising statistical approach to analyse this type of data is multilevel two-part modelling, which accounts for the potential dependency between the occurrence and the amount of food intake (see Ruf, Neubauer, et al, 2021, for a detailed description of this approach).…”