Fig 1.
Histograms of predicted participation probabilities at baseline (a), the 2005 follow-up (b), and the 2010 follow-up (c). Predictions were obtained from background characteristics (see Table 1 as well as Tables A1 and A2 in S1 File) and are plotted separately for participants and non-participants of the survey in question. Non-participants in panels (b) and (c) only include individuals who had participated in the previous round(s). The bar width was set to 0.05.
Table 1.
Background characteristics that were used to estimate baseline participation probability.
Table 2.
Unweighted and reweighted associations (multivariable).
Fig 2.
Histogram of relative changes in (multivariable) associations.
The figure shows ratios of log odds ratios from reweighted and unweighted multivariable models. The bar width was set to 0.1.
Fig 3.
Directed acyclic graphs (DAGs) showing how selection bias can arise.
Panel A displays an example of collider bias (where estimates from the sample lack causal interpretation), whereas panel B displays an example of causal effect heterogeneity (where the effects are inherently different across participants and the full population).