Supervisors info:
Νάσκα Ανδρονίκη, Αναπληρώτρια Καθηγήτρια,Ιατρικής Στατιστικής, ΕΚΠΑ
Μπάμια Χριστίνα, Αναπληρώτρια Καθηγήτρια, Εργαστήριο Υγιεινής, Επιδημιολογίας και Ιατρικής Στατιστικής, ΕΚΠΑ
Ορφανός Φίλιππος, Ειδικό και Εργαστηριακό Διδακτικό Προσωπικό με γνωστικό αντικείμενο «Ιατρική Στατιστική - Επιδημιολογία», Εργαστήριο Υγιεινής, Επιδημιολογίας και Ιατρικής Στατιστικής, ΕΚΠΑ
Summary:
Abstract
Introduction: The interest for epidemiological research on diet-health relation has been considerably increased during the last few decades. The usual dietary intake estimation both on population and individual level is therefore essential. However, several traits of the dietary intake information necessitate the development and use of specialized statistical models allowing for usual dietary intake estimation with the least possible error.
Aim: The primary aim of this thesis is to compare the usual dietary intake estimations produced for energy and specific food components (red meat, pulses, fruits and vegetables, sugar and confectionary, lipids, iron and vitamin C), derived naively (mean of two 24 hour dietary recalls) and via variants of the method developed by National Cancer Institute (NCI method). The secondary aim is to compare the estimated intakes with the current dietary guidelines. Essentially, the NCI method is the application of specifically designed statistical models accounting for dietary measurement error.
Methodology: The dietary data as well as other personal information that was used in this thesis originate from the preliminary database of the study "HYDRIA-Program and focused action for health and nutrition of the Greek population: methodology development and application and mapping" which was conducted by the Hellenic Health Foundation in collaboration with the Center for disease control and prevention (ΚΕ.ΕΛ.Π.ΝΟ) of the Ministry of Health. The dietary data analysis consists of a series of NCI method's variants, followed by systematic comparison between them, including the naive/plain usual intake estimation (mean of two 24 hours dietary recalls per person) as one of the compared approaches. The NCI method's variants differ as to the way the consumption is being modeled (modeling of quantity of consumption versus modeling of both possibility and quantity of consumption), the relation between the random effects components of the model (correlated versus not correlated), the use versus not use of covariates (age, gender, education level) and the use versus not use of additional dietary information provided by a qualitative food frequency questionnaire. The estimated usual intake distributions derived by the various approaches are compared in pairs via the Harrel and Davis method for distributional comparisons with use of bootstrap estimators for the percentiles of consumption. The estimated distributions are matched in pairs, in a suitable way that enables the direct evaluation of each part that differs between the various models. Finally, the estimated percentage of non compliance with current dietary guidelines (World Health Organization and European Food Safety Authority) as derived by the plain approach and the NCI method's variants are compared.
Results-Conclusions: Different approaches selected for the analysis of a dietary component can potentially lead to different estimations of the usual intake distribution of the population. The plain approach seems to be reliable when estimating the mean usual intake of a dietary component in the population. It tends however to underestimate usual intake in low percentiles of distribution of usual intake and overestimate usual intake at the higher percentiles. For daily or nearly daily consumed dietary components the NCI amount model is the recommended choice. For episodically consumed dietary components the NCI two-part model, with correlated or non correlated person specific (random) effects is recommended. The inclusion of supportive covariates in the model does not affect the estimation of the usual intake distribution of the population. The main use of these covariates is to help estimate the usual intake distribution of various subpopulations. The dietary information of the FFQ seems to lead to statistically significant changes when included in the NCI two part model with non correlated person specific effects but not when included in the NCI two part model with correlated person specific effects. However, in the case of pulses (89% zero reported consumption in the recalls), when FFQ information is included in the correlated two part model, the distribution's estimation changes considerably, especially in the extreme low and high percentiles. However, these changes are not statistically important, maybe because of the immense uncertainty that is introduced due to the zero inflated data. The non-compliance with the dietary guidelines percentage tends to be either overestimated or underestimated when the plain approach (mean of two 24 hours dietary recalls per person) is used. However, the estimated non compliance percentage in the population as derived by the NCI method's variants, do not differ between the recommended, for each dietary component, variants.
Keywords:
NCI method application, Comparison of usual dietary intake distributions, Compliance with dietary guidelines, Bootstrap methodology, Measurement error in dietary data
