Variable Selection

"For Phase I, the following step-wise modelling approach was conducted: (i) each acute-phase symptom that occurred in more than 2% of the patients, was analyzed in a univariable logistic regression model; (ii) symptoms that showed a two-sided p-value smaller than 0.05 in (i) were analyzed within the respective system-specific multivariable model (neurological, cardiovascular, respiratory, gastrointestinal, musculoskeletal, and psychiatric) with backward likelihood-ratio elimination using a removal criterion of p < 0.05; (iii) All remaining parameters identified in (ii) were entered in an overall multivariable model together with further clinical relevant parameters (age-group: < 50 years vs. ≥ 50 years, biological sex: male vs. female, and obesity: BMI < 30 kg/m2 vs. ≥ 30 kg/m2) and again selected with a backward likelihood-ratio selection. Given the large number of candidate symptoms, a stepwise modelling strategy with intermediate system-specific models was used to limit model complexity, while recognizing the known limitations of data-driven variable selection16. A removal criterion of p < 0.05 was applied consistently throughout the backward selection procedures.\n\nThe same step-wise modelling approach was applied in Phase II. For participants who reported reinfection between the two surveys, acute-phase symptoms referred to the most recent SARS-CoV-2 infection, whereas for participants without reinfection, acute-phase symptoms referred to the index infection. In an additional analysis addressing the influence of variable selection, the Phase II model was re-estimated using the covariate set entered into the combined Phase I model, applying the same backward likelihood-ratio selection procedure. In a further sensitivity analysis, the final Phase II model was additionally adjusted for reinfection status to assess whether the observed associations were materially influenced by reinfection between Phase I and Phase II. As sensitivity analyses of the Phase I and Phase II PCC models, unselected multivariable logistic regression models were fitted by entering all eligible acute-phase symptoms with a prevalence > 2% simultaneously together with age group, sex, and obesity, without p-value-based variable selection."