Given the risk of selection bias in circumstances where underlying missingness mechanisms are informative, a chained equations imputation model was created under the assumption that missing drinking data were predictable from observed covariates 20, 21. Using the ‐mi‐ package 22, missing data for participants who were lost to follow‐up or else provided no response to the alcohol consumption questions of interest were predicted from a range of demographic, socio‐economic, health and life‐style characteristics. A total of 50 imputations were run to be sure of capturing appropriately the degree of uncertainty surrounding the predicted values, with iterations run for each imputation until predicted values reached convergence. The imputation model excluded missing data for phases of observation on or after any documented date of death. Finally, to ensure that the estimation sample was consistent between imputations, baseline consumption categories were defined using the observed data only, with imputed volumes of alcohol consumption then predicted for all follow‐up phases.
