Fig. 4: Modelling of persistent symptoms as a function of biological and demographic variables.

a Logistic regression models with one or more symptoms at 12 weeks (y/n) as the binary outcome variable, both adjusted for age-sex and mutually adjusted*; b mean contribution to area under the curve (AUC) that each variable makes to a multivariable boosted tree model, derived by permuting each variable in turn (1000× to obtain a distribution) and quantifying the change in model performance; c modelled probability of persistent symptoms at 12 weeks as a function of age and sex, using generalised additive models with splines on age and interactions between age and sex. All models were fit on n = 71,642 respondents for whom we had 150 days’ observation time. Age, sex, adiposity household income, healthcare/care home worker, deprivation, current smoker status and prior hospitalisation with COVID-19 are the strongest predictors of persistent symptoms in multivariable modelling, while Asian ethnicity is associated with a lower risk of persistent symptoms at 12 weeks. Box plots in panel b show median, first and third quartiles; whiskers indicate 1.5 × the interquartile range; data beyond this range are plotted as points. Note: Owing to missing data in some variables, the total n for the mutually adjusted model in panel a is 55,730.