Methodological issues in measuring alcohol consumption

One important goal of alcohol epidemiology is to link alcohol consumption with alcohol-related problems. To this end, alcohol consumption first must be determined as accurately as possible. Surveys of consumption allow researchers to ask individuals about their drinking patterns and various methodological issues influence the measurement of alcohol consumption from these datasets. Two commonly used measures are the usual quantity/frequency and graduated frequency approaches, both of which allow researchers to estimate the volume of alcohol intake. These methods mainly rely on the assumption that distribution of observation within each category is uniform, leading to histograms with wide bins.

In the last year, I have started to collaborate with David Leon from the London School Of Hygiene and Tropical Medicine on techniques to reconstruct, from ordinal data, a continuous latent distribution of alcohol consumption. Respondents in a survey tick a specific category, and this observed frequency will produce a so-called probability mass function. This situation can be viewed as grouped data into coarse intervals. However, the underlying distribution can not be scattered: we aim to estimate a smooth curve under the condition that integrals calculated over the intervals of a given width are equal to the observed counts.