Repository logo
 

Multivariate Analysis of Dietary Data in the Presence of Excess Zeros and Measurement Error


Type

Thesis

Change log

Authors

Chernova-Chernaya, Yulia 

Abstract

Nutritional epidemiology is a complex area of research plagued with bias and measurement error partly arising from the use of imperfect methods of dietary data measurement and inadequate data analyses. Some foods or nutrients are consumed habitually, while others are consumed occasionally. This thesis addresses three important public health problems: (1) Estimation of under and over consumption of foods consumed occasionally such as alcohol; (2) Investigation of the determinants of food intake, and (3) Estimation of the effect of the intake of occasionally-consumed foods on health outcomes.

The analysis of food intake is complicated because foods are often eaten in combination. This gives rise to multiple correlated habitually- and occasionally-consumed food intakes, characterised by a large proportion of zero observations. Modern statistical methods are required to deal with measurement error, excess zeros in the intake distributions, and correlated preferences for frequency of consumption and portion sizes across foods.

The thesis demonstrates the use of contemporary statistical methods, based on mixed-distributions and mixed-effects modelling approaches, for the analysis of a single and multiple correlated habitually and occasionally-consumed food intakes. These methods are complex due to the need to evaluate intractable integrals for parameter estimation.

Firstly, to describe under and over consumption, the thesis provides a new numerical approach, which is a quicker and simpler alternative to Monte Carlo simulation, to estimate the distributional quantiles of occasionally-consumed food intakes in predefined sub-populations.

Secondly, dietary data from the UK National Diet and Nutrition Survey Rolling Programme (NDNS RP), which provides the only source of current authoritative information on food and nutrient intake in the UK, are analysed with a mixed-effects two-part model to estimate the associations between personal and socio-economic risk factors and the intake of several foods of current public health importance. %

This is the first time this approach is applied to NDNS RP data to assess explicitly socio-economic and personal characteristics related to occasionally-consumed food intakes in the UK population.

Then, the thesis develops a novel multivariate joint model for several correlated occasionally-consumed food intakes utilising a pseudolikelihood approach and parametric bootstrap for parameter estimation. The approach is illustrated by modelling the intake of alcohol, jointly with the intake of other foods, and the resulting analysis is compared with that based on the two-part model for a single food and the traditional multivariable linear regression model widely used in nutritional epidemiology.

Finally, a regression calibration approach is applied to correct for the effect of excess zeros and multiple correlated person-specific preferences when several food intakes are investigated as predictors of health outcomes. Again the results are contrasted with those obtained by applying multivariable regression analysis which ignores excess zeros and correlated preferences, introducing potential bias in the effect estimates. As an example, the relationships between alcohol intake in a male sub-population of NDNS RP and haemoglobin A1c (HbA1c), a known predictor of type 2 diabetes mellitus, are investigated. Ideally, to obtain unbiased estimates of predictors' effects, all correlated unobserved person-specific effects should be accounted for. However, the task is incredibly complex and to tackle this the suggestion is to simplify the model by accounting only for the largest residual correlations. Even in this imperfect form, regression calibration produces markedly different results from multivariate linear regression.

Description

Date

2019-07-01

Advisors

Solis-Trapala, Ivonne
Wood, Angela

Keywords

semi-continuous outcome, measurement error, excess zeros, occasionally-consumed food, regression calibration

Qualification

Doctor of Philosophy (PhD)

Awarding Institution

University of Cambridge