Julia Polak

What is the recipe for a long life: good diet, exercise, both? Statistical methods for assessing the causal effect of dynamic interventions.

Julia Polak1 joint work with Lyle Gurrin1, Elizabeth Williamson2, Julie Simpson1, Allison Hodge3, Julie Basset3 and Michael Fahey2, Graham Giles3, Andrew Forbes2 and Dallas English1

1 Melbourne School of Population Health, The University of Melbourne.
2 Department of Epidemiology and Preventive Medicine, Monash University.
3 Cancer Epidemiology Centre, Cancer Council Victoria.

 

Abstract: Poor diet is an acknowledged risk factor for life-threatening conditions such as cancer and cardiovascular disease. The association between improving diet and reducing mortality risk has, however, received little attention. Changes in diet are likely to be associated with other lifestyle choices, therefore may influence or be influenced by (i.e. confounded) other risk factors.

To study the effect of a dietary regimen we need statistical methods that provide unbiased estimates of parameters that have a causal interpretation in the presence of time-dependent confounding. One such method is g-computation, a type of standardisation procedure that allows us to contrast the marginal mean risk between dietary regimens at the population level while adjusting for time varying confounders such as physical activity, alcohol consumption and total energy intake. The parametric g-formula is a computationally simpler approach to g-computation, which otherwise becomes unwieldy in the presence of multiple covariates and dynamic interventions.

This removes the need to evaluate high-dimensional integrals by assuming distributional forms for the relationship between the outcome and covariates, and current and previous covariate profile. A system of equations is used to predict the mortality risk on comparative dietary regimens where covariate values of participants are "reset" continually to ensure they are consistent with the proposed intervention, such as improving diet in combination with increasing physical activity and reducing alcohol consumption, for some or all individuals at some or all times during follow-up.

We demonstrate the performance of the parametric g-formula on the Melbourne Collaborative Cohort Study (MCCS) data set, which includes 41,000 participants assessed on 3 occasions over 15 years. In addition we calculate a Mediterranean Diet Score (MDS) from MCCS participants' responses to a Food Frequency Questionnaire and report the results on the causal effect of maintaining the MDS in the "healthy" range for mortality risk.

Keywords: Chow test, Prediction capability testing, P value, Multivariate kernel density estimation, Structural break.

 

 

 








 

 

 

 




 

 

PhD Student

             Department of Econometric & Business Statistics

                          Monash University

CV